JP2015129987A - System and method of forming medical high-resolution image - Google Patents
System and method of forming medical high-resolution image Download PDFInfo
- Publication number
- JP2015129987A JP2015129987A JP2014000148A JP2014000148A JP2015129987A JP 2015129987 A JP2015129987 A JP 2015129987A JP 2014000148 A JP2014000148 A JP 2014000148A JP 2014000148 A JP2014000148 A JP 2014000148A JP 2015129987 A JP2015129987 A JP 2015129987A
- Authority
- JP
- Japan
- Prior art keywords
- resolution
- image
- learning
- low
- patches
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 43
- 238000002591 computed tomography Methods 0.000 claims description 17
- 238000011156 evaluation Methods 0.000 claims description 12
- 238000002595 magnetic resonance imaging Methods 0.000 claims description 6
- 230000015572 biosynthetic process Effects 0.000 claims description 4
- 238000005516 engineering process Methods 0.000 abstract description 14
- 238000012545 processing Methods 0.000 abstract description 12
- 230000006872 improvement Effects 0.000 abstract description 4
- 238000012805 post-processing Methods 0.000 abstract description 2
- 238000012615 high-resolution technique Methods 0.000 description 26
- 210000004351 coronary vessel Anatomy 0.000 description 14
- 238000003384 imaging method Methods 0.000 description 8
- 238000002474 experimental method Methods 0.000 description 5
- 230000006870 function Effects 0.000 description 5
- 230000008569 process Effects 0.000 description 5
- 238000005070 sampling Methods 0.000 description 5
- 238000003745 diagnosis Methods 0.000 description 4
- 238000010586 diagram Methods 0.000 description 4
- 230000001629 suppression Effects 0.000 description 4
- 238000009499 grossing Methods 0.000 description 3
- KUVIULQEHSCUHY-XYWKZLDCSA-N Beclometasone Chemical compound C1CC2=CC(=O)C=C[C@]2(C)[C@]2(Cl)[C@@H]1[C@@H]1C[C@H](C)[C@@](C(=O)COC(=O)CC)(OC(=O)CC)[C@@]1(C)C[C@@H]2O KUVIULQEHSCUHY-XYWKZLDCSA-N 0.000 description 2
- 201000000057 Coronary Stenosis Diseases 0.000 description 2
- 206010011089 Coronary artery stenosis Diseases 0.000 description 2
- 208000031481 Pathologic Constriction Diseases 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 2
- 239000008280 blood Substances 0.000 description 2
- 210000004369 blood Anatomy 0.000 description 2
- 210000004204 blood vessel Anatomy 0.000 description 2
- 230000015556 catabolic process Effects 0.000 description 2
- 238000010276 construction Methods 0.000 description 2
- 238000006731 degradation reaction Methods 0.000 description 2
- 238000010988 intraclass correlation coefficient Methods 0.000 description 2
- 230000002107 myocardial effect Effects 0.000 description 2
- 230000009467 reduction Effects 0.000 description 2
- 230000000241 respiratory effect Effects 0.000 description 2
- 101150008563 spir gene Proteins 0.000 description 2
- 230000036262 stenosis Effects 0.000 description 2
- 208000037804 stenosis Diseases 0.000 description 2
- 206010028980 Neoplasm Diseases 0.000 description 1
- 238000002583 angiography Methods 0.000 description 1
- 210000000988 bone and bone Anatomy 0.000 description 1
- 210000004556 brain Anatomy 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 239000002131 composite material Substances 0.000 description 1
- 238000013170 computed tomography imaging Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000002059 diagnostic imaging Methods 0.000 description 1
- 229940079593 drug Drugs 0.000 description 1
- 239000003814 drug Substances 0.000 description 1
- 230000003902 lesion Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 210000000056 organ Anatomy 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 239000013598 vector Substances 0.000 description 1
Abstract
Description
本手法は、コンピュータ断層撮影（CT: Computer Tomography）画像や磁気共鳴(MR:
Magnetic Resonance)画像などの医用画像を高解像化、高画質化するための医用高解像画像形成システムである。
This technique is based on computed tomography (CT) images and magnetic resonance (MR).
This is a high-resolution medical image forming system for improving the resolution and image quality of medical images such as magnetic resonance images.
近年、CT撮影装置やMR撮影装置などの医用画像撮影装置が著しく進歩し、病変の早期発見に大きく貢献している。しかし、それらの装置で得られた医用画像の画質、解像度がしばしば十分でない場合があり、その高画質化、高解像化が望まれている。これを解決する手法の一つとして、観測された低画質画像または低解像画像（低画質／低解像画像）から高画質画像または高解像画像（高画質／高解像画像）を推定する超解像技術に関する研究が行われている。 In recent years, medical imaging devices such as CT imaging devices and MR imaging devices have made significant progress and have greatly contributed to early detection of lesions. However, there are cases where the quality and resolution of medical images obtained with these devices are often not sufficient, and higher image quality and higher resolution are desired. One method to solve this is to estimate high-quality images or high-resolution images (high-quality / high-resolution images) from observed low-quality images or low-resolution images (low-quality / low-resolution images). Research on super-resolution technology is underway.
超解像技術は「再構成型」と「学習型」の大きく２つに分けられる。観測された低画質／低解像画像ｙは、高画質／高解像画像ｘに歪み関数、広がり関数、縮小関数を乗算し、更にノイズを加算したものとして定義される。再構成型の超解像技術では、まず、適当な初期高画質／高解像画像ｘ０を与え、それに対する低画質／低解像画像ｙ０を計算する。そして、ｙｉと観測された低画質／低解像画像ｙの誤差が小さくなるようにｘｉを繰り返し更新することにより、高画質／高解像画像を推定する。 Super-resolution technology can be broadly divided into “reconstruction type” and “learning type”. The observed low-quality / low-resolution image y is defined as a high-quality / high-resolution image x multiplied by a distortion function, a spread function, and a reduction function, and further added with noise. In the reconstruction-type super-resolution technique, first, an appropriate initial high-quality / high-resolution image x0 is given, and a low-quality / low-resolution image y0 corresponding thereto is calculated. Then, the high image quality / high resolution image is estimated by repeatedly updating xi so that the error of the low image quality / low resolution image y observed as yi becomes small.
再構成型は汎用性が高く、さまざまな画像に適用できる一方、画質の大幅な改善、解像度の大幅な向上は望めないことが知られている。学習型の超解像技術は、低画質／低解像画像と高画質／高解像画像のペアで構成される学習用画像を用いて、その対応関係を学習し、入力された低画質／低解像画像に対する高画質／高解像画像を推定する。学習型は適用分野を限定し、適切な学習用画像を用いることにより、画質の大幅な改善、解像度の大幅な向上が期待できる。超解像技術の医療分野の応用として、再構成型、学習型ともに脳MRへ適用した研究が数多く報告されている。 It is known that the reconstruction type has high versatility and can be applied to various images, but a great improvement in image quality and a great improvement in resolution cannot be expected. The learning-type super-resolution technology uses a learning image composed of a pair of a low image quality / low resolution image and a high image quality / high resolution image to learn the correspondence, and the input low image quality / Estimate a high quality / high resolution image for a low resolution image. The learning type limits application fields, and by using an appropriate learning image, it can be expected to greatly improve image quality and resolution. There have been many reports on the application of super-resolution technology to brain MR for both reconstruction and learning.
一般に、医用画像撮影装置から得られた医用画像の画質、解像度が高いほど、医師の診断能が向上することが知られている。多くの医療機器メーカーが撮影装置を改良し、定期的に新製品を発表しているが、買い替え費用が大きな問題となり、短期間に普及しないのが現状である。そこで、観測された医用画像を後処理としてソフトウェア的に高画質化、高解像化する画像処理手法が望まれている。本発明が解決しようとする課題は、学習型の超解像技術を用いて、医用画像の高解像化、高画質化のための医用高解像画像形成システムを提供することである。 In general, it is known that the diagnostic ability of a doctor improves as the image quality and resolution of a medical image obtained from a medical image photographing apparatus increases. Many medical device manufacturers improve their imaging devices and regularly announce new products, but the cost of replacement is a major problem and is not widespread in a short time. Therefore, an image processing technique for improving the image quality and resolution in software as post-processing of the observed medical image is desired. The problem to be solved by the present invention is to provide a medical high-resolution image forming system for improving the resolution and image quality of medical images using a learning type super-resolution technique.
請求項１に記載の発明は、コンピュータ断層撮影（ＣＴ）システム及び、磁気共鳴イメージング（ＭＲＩ）システムで作成される低解像の医用画像を入力画像とし、入力画像を高解像出力画像にして出力する医用高解像画像形成システムである。
本発明の医用高解像画像形成システムは、少なくとも学習用高解像画像入力部と、辞書作成部と、辞書記憶部と画像入力部と、高解像出力画像作成部を有する。
学習用高解像画像入力部は、１以上の複数の学習用高解像画像が辞書作成のため入力される。入力された各学習用高解像画像は、辞書作成部で画素数がＭ分の１の縮小画像に変換されて学習用低解像画像が作成される。
According to the first aspect of the present invention, a low-resolution medical image created by a computed tomography (CT) system and a magnetic resonance imaging (MRI) system is used as an input image, and the input image is used as a high-resolution output image. It is a medical high-resolution image forming system for outputting.
The medical high-resolution image forming system of the present invention includes at least a learning high-resolution image input unit, a dictionary creation unit, a dictionary storage unit, an image input unit, and a high-resolution output image creation unit.
The learning high-resolution image input unit receives one or more learning high-resolution images for dictionary creation. Each input high resolution image for learning is converted into a reduced image having a number of pixels of 1 / M by a dictionary creation unit to create a low resolution image for learning.
また、この学習用低解像画像はＮ×Ｎサイズの学習用低解像画像パッチに分割され、この学習用低解像画像パッチに対応させて学習用高解像画像もＭＮ×ＭＮサイズの学習用高解像画像パッチに分割される。
また、学習用低解像画像パッチと学習用高解像画像パッチは対応付けされて辞書記憶部格納される。
The learning low-resolution image is divided into N × N-size learning low-resolution image patches, and the learning high-resolution image corresponding to the learning low-resolution image patch is also MN × MN-sized. Divided into learning high-resolution image patches.
Further, the learning low-resolution image patch and the learning high-resolution image patch are associated with each other and stored in the dictionary storage unit.
一方、画像入力部には低解像画像である入力画像が入力され、高解像出力画像作成部において入力画像がＮ×Ｎサイズの入力画像パッチに分割され、分割された各入力画像パッチは辞書記憶部に格納された学習用低解像画像パッチの全てと個別に対比され類似性評価が行われる。この類似性評価において、類似性評価値の高い順にＫ個の学習用低解像画像パッチが抽出される。次いで、抽出されたＫ個の前記学習用低解像パッチに対応づけされた前記学習用高解像パッチが抽出され、これら抽出されたＫ個の学習用高解像パッチを結合して高解像出力画像を作成し、高解像出力画像として出力する。 On the other hand, an input image which is a low resolution image is input to the image input unit, and the input image is divided into N × N size input image patches in the high resolution output image creation unit, and each divided input image patch is Similarity evaluation is performed by individually comparing all of the learning low-resolution image patches stored in the dictionary storage unit. In this similarity evaluation, K learning low-resolution image patches are extracted in descending order of the similarity evaluation value. Next, the learning high resolution patches associated with the extracted K learning low resolution patches are extracted, and the extracted K learning high resolution patches are combined to obtain a high resolution. An image output image is created and output as a high resolution output image.
請求項２に記載の発明は、請求項１に記載の医用高解像画像形成システムであって、前記類似性評価を入力画像パッチ内の画素データ値と、辞書記憶部に記憶された学習用低解像パッチの画素データ値との特徴量空間におけるユークリッドの距離に基づいて類似性評価値を定める。これによって、入力画像パッチに類似する学習用低解像パッチを一義的に抽出することができる。 A second aspect of the present invention is the medical high-resolution image forming system according to the first aspect, wherein the similarity evaluation is performed for pixel data values in an input image patch and for learning stored in a dictionary storage unit. The similarity evaluation value is determined based on the Euclidean distance in the feature amount space with the pixel data value of the low resolution patch. As a result, a learning low-resolution patch similar to the input image patch can be uniquely extracted.
請求項３に記載の発明は、請求項１または請求項２に記載の医用高解像画像形成システムであって、ＣＴまたは、ＭＲＩで撮像される複数の撮像画像が全て入力画像とされて高解像出力画像が作成される。これら複数の撮像画像は断層写真であり、高解像出力画像は３次元画像として出力でき、３次元医用高解像画像形成システムの機能も備える。 A third aspect of the present invention is the medical high-resolution image forming system according to the first or second aspect, wherein a plurality of captured images captured by CT or MRI are all set as input images. A resolution output image is created. The plurality of captured images are tomographic images, and the high-resolution output image can be output as a three-dimensional image, and has a function of a three-dimensional medical high-resolution image forming system.
請求項４に記載の医用高解像画像形成方法は、コンピュータ断層撮影（ＣＴ）システム及び、磁気共鳴イメージング（ＭＲＩ）システムで作成される低解像の医用画像を入力画像とし、高解像出力画像に変換して出力する医用高解像画像形成方法である。
本発明の医用高解像画像形成方法は、少なくとも１つの学習用高解像画像の画素数をＭ分の１に縮小した学習用低解像画像を作成する。
次いで、学習用低解像画像をＮ×Ｎサイズの学習用低解像画像パッチに分割し、これら学習用低解像画像パッチに対応させて学習用高解像画像をＭＮ×ＭＮサイズの学習用高解像画像パッチに分割する。
5. The medical high-resolution image forming method according to claim 4, wherein a low-resolution medical image created by a computed tomography (CT) system and a magnetic resonance imaging (MRI) system is used as an input image, and a high-resolution output is provided. This is a medical high-resolution image forming method for converting to an image and outputting it.
The medical high-resolution image forming method of the present invention creates a learning low-resolution image in which the number of pixels of at least one learning high-resolution image is reduced to 1 / M.
Next, the learning low-resolution image is divided into N × N-size learning low-resolution image patches, and the learning high-resolution image corresponding to these learning low-resolution image patches is learned in MN × MN size. Divide into high-resolution image patches.
また学習用低解像画像パッチと前記学習用高解像画像パッチは、対応付けされた辞書として、辞書記憶部に格納される。
一方、画像入力部から入力される低解像の入力画像は、Ｎ×Ｎサイズの入力画像パッチに分割され、分割された各入力画像パッチは、辞書記憶部に格納された学習用低解像画像パッチの全てと個別に対比させられ類似性が評価される。また、類似性評価値の高い順にＫ個の学習用低解像画像パッチが抽出され、抽出されたＫ個の学習用低解像パッチに対応づけされた学習用高解像パッチが抽出される。次いで、抽出されたＫ個の学習用高解像パッチを合成して高解像出力画像を作成する。全ての入力画像パッチについてこの作業を実施し、最終的に高解像出力画像を得る医用高解像画像形成方法の発明である。
The learning low-resolution image patch and the learning high-resolution image patch are stored in the dictionary storage unit as an associated dictionary.
On the other hand, the low resolution input image input from the image input unit is divided into N × N size input image patches, and each divided input image patch is stored in the dictionary storage unit for learning low resolution. Each image patch is individually compared and the similarity is evaluated. Further, K learning low-resolution image patches are extracted in descending order of similarity evaluation value, and learning high-resolution patches associated with the extracted K learning low-resolution patches are extracted. . Next, the extracted K high-resolution patches for learning are synthesized to create a high-resolution output image. This invention is an invention of a medical high-resolution image forming method in which this operation is performed for all input image patches and a high-resolution output image is finally obtained.
上記医用高解像画像形成システムをＣＴ画像やＭＲ画像などに適用することにより、撮影装置の撮像センサーの解像度に起因する低解像画像で失われている画素間の細かな高周波成分を推定した高解像画像を生成することが可能である。また、撮像方法自体に起因する画質劣化や、心拍動など身体の動きに起因する画質劣化に伴う低画質画像をぼけやノイズを抑制した高画質画像へ改善することが可能である。 By applying the above medical high-resolution image forming system to CT images and MR images, fine high-frequency components between pixels lost in the low-resolution image due to the resolution of the imaging sensor of the imaging device were estimated. It is possible to generate a high resolution image. In addition, it is possible to improve a low-quality image due to image quality degradation caused by the imaging method itself or image quality degradation caused by body movement such as heartbeat to a high-quality image with reduced blur and noise.
本発明による医用高解像画像形成システムの構成を図１に、医用高解像画像形成システムの機能説明を図２に示す。医用高解像画像形成システム１０は、学習用高解像画像入力部１１と、辞書作成部１２と、辞書記憶部１３と、画像入力部１４と、高解像出力画像作成部１５で構成される。
ＣＴ画像やＭＲ画像などはボリュームデータであり、多くのスライス画像で構成される。したがって、入力画像からＭ倍の解像度を有する高解像画像を推定する場合には、入力画像を含めた同じ患者のスライス画像を学習用高解像画像１６として用い、それらを１／Ｍのサイズにダウンサンプリングすることにより学習用低解像画像２１を生成する。
FIG. 1 shows a configuration of a medical high-resolution image forming system according to the present invention, and FIG. 2 shows a functional description of the medical high-resolution image forming system. The medical high-resolution image forming system 10 includes a learning high-resolution image input unit 11, a dictionary creation unit 12, a dictionary storage unit 13, an image input unit 14, and a high-resolution output image creation unit 15. The
CT images, MR images, and the like are volume data and are composed of many slice images. Therefore, when estimating a high-resolution image having a resolution of M times from the input image, the slice image of the same patient including the input image is used as the learning high-resolution image 16, and the size is 1 / M. The learning low resolution image 21 is generated by down-sampling.
そして、それらの学習用高解像画像１６、学習用低解像画像２１のそれぞれから切り出したMN×MN(×MN)サイズの学習用高解像パッチ２４とN×N(×N)サイズの学習用低解像パッチ２３のペアで辞書２２を構築する。次に、入力画像１７を各画素でN×N(×N)サイズのパッチに切り出し、各入力パッチに対して辞書２２から特徴量空間のユークリッドの距離が最小となる学習用低解像パッチを類似パッチとしてＫ個選択する。そして、類似パッチの線形和で入力パッチを復元するための重み係数を最小二乗法で計算し、その重み係数と類似パッチに対応する学習用高解像パッチ２４の線形和で、入力パッチ２５に対する高解像パッチ１８を生成する。 Then, a learning high resolution patch 24 of MN × MN (× MN) size cut out from each of the learning high resolution image 16 and learning low resolution image 21 and an N × N (× N) size. A dictionary 22 is constructed with a pair of learning low-resolution patches 23. Next, the input image 17 is cut into N × N (× N) size patches at each pixel, and a learning low-resolution patch that minimizes the Euclidean distance in the feature amount space from the dictionary 22 is input to each input patch. Select K similar patches. Then, a weighting factor for restoring the input patch by a linear sum of similar patches is calculated by the least square method, and the weighting factor and a linear sum of the learning high-resolution patches 24 corresponding to the similar patches are used for the input patch 25. A high resolution patch 18 is generated.
この処理をすべての入力パッチ２５に適用し、高解像パッチが重複する領域はその平均値を求めることにより、Ｍ倍の高解像画像１８を生成する。ただし、骨組織などは高解像化により、医師が異質と違和感を覚える可能性があるため、必要に応じ、ＣＴ値、信号強度、テクスチャ情報や位置情報に基づいて、高解像化の対象から除外する。 This process is applied to all the input patches 25, and an M-fold high-resolution image 18 is generated by obtaining an average value of the regions where the high-resolution patches overlap. However, because of the high resolution of bone tissue, etc., doctors may feel dissimilar and uncomfortable, so if necessary, based on the CT value, signal strength, texture information and position information, Exclude from
次に本発明について更に詳細に説明を行う。
ＣＴ画像やＭＲ画像はボリュームデータであり、多くのスライス画像で構成される。したがって、入力画像からＭ倍の解像度を有する高解像画像を推定する場合には、入力画像を含めた同じ患者のスライス画像を学習用高解像画像として用い、それらを標準偏差Ｍ／２画素のガウシアン平滑化フィルタを適用した後、１／Ｍのサイズにダウンサンプリングすることにより学習用低解像画像を生成する。ただし、対象とする医用画像がボリュームデータでない場合など同じ患者のスライス画像を使用できない場合には、２０名程度の別患者の画像を学習用画像として使用する。そして、学習用低解像画像から切り出したN×N(×N)サイズの学習用低解像パッチと、それに対応する位置にある学習用高解像画像のMN×MN(×MN)サイズの学習用高解像パッチをペアにし、それを辞書に格納する。
Next, the present invention will be described in more detail.
CT images and MR images are volume data and are composed of many slice images. Therefore, when estimating a high-resolution image having a resolution of M times from the input image, the slice image of the same patient including the input image is used as a high-resolution image for learning, and these are used as standard deviation M / 2 pixels. After applying the Gaussian smoothing filter, a low resolution image for learning is generated by down-sampling to 1 / M size. However, when slice images of the same patient cannot be used, such as when the target medical image is not volume data, images of about 20 other patients are used as learning images. Then, a learning low resolution patch of N × N (× N) size cut out from the learning low resolution image and the MN × MN (× MN) size of the learning high resolution image at the corresponding position Pair the learning high-resolution patches and store them in the dictionary.
辞書の構築において、様々な入力パッチに対応できるよう各学習用パッチを９０度、１８０度、２７０度回転し４ペアを生成する。図３学習用低解像パッチ２４と、これとペアになる学習用高解像パッチ２５の例を示す。ただし、すべてのパッチペアを辞書に格納するとその数が数百万〜数千万となり、演算時間が膨大となることから、類似した学習用パッチは削除する。また、N×N(×N)のパッチサイズは、入力画像のサイズ、ピクセルサイズ、診断対象（例えば、臓器、血管、腫瘍など）の大きさに応じて、適切に与える。 In the construction of the dictionary, each learning patch is rotated 90 degrees, 180 degrees, and 270 degrees so as to correspond to various input patches to generate four pairs. FIG. 3 shows an example of a learning low resolution patch 24 and a learning high resolution patch 25 paired therewith. However, if all the patch pairs are stored in the dictionary, the number thereof becomes several million to several tens of millions, and the calculation time becomes enormous. Therefore, similar learning patches are deleted. The patch size of N × N (× N) is appropriately given according to the size of the input image, the pixel size, and the size of the diagnosis target (for example, organ, blood vessel, tumor, etc.).
次に、高解像化アルゴリズムにおいて、入力画像を各画素でN×N(×N)サイズのパッチに切り出し、各入力パッチｙＬに対して辞書から特徴量空間のユークリッドの距離が最小となる学習用低画質／低解像パッチを類似パッチｘＬ ｉとしてＫ個選択する。
図４にＫ＝７の類似パッチの選択例を示す。この特徴量空間は、N×N(×N)サイズの各パッチの１次差分値、２次差分値、平均値などのベクトルで構成され、診断対象に応じて、その特徴量を適切に選択する。例えば、ＭＲ画像のボリュームデータに含まれる血管が診断対象で、その境界を明瞭にしたい場合には、パッチサイズを５×５×５で与え、移動平均値とｘ方向、ｙ方向、ｚ方向の各１次差分値により５×５×５×４次元の特徴量空間で構成される。そして、類似パッチＫ個の線形和（式１）で入力パッチを復元するための重み係数を最小二乗法（式２）で計算する。
FIG. 4 shows an example of selecting similar patches with K = 7. This feature amount space is composed of vectors such as the primary difference value, secondary difference value, and average value of each patch of N × N (× N) size, and the feature amount is appropriately selected according to the diagnosis target To do. For example, when a blood vessel included in the volume data of the MR image is a diagnosis target and it is desired to make the boundary clear, the patch size is given by 5 × 5 × 5, and the moving average value and the x, y, and z directions are given. Each primary difference value is composed of a 5 × 5 × 5 × 4 dimensional feature amount space. Then, a weighting factor for restoring the input patch with a linear sum of K similar patches (Equation 1) is calculated by the least square method (Equation 2).
その重み係数と類似パッチに対応する学習用高解像パッチｘＨ ｉの線形和（式３）で、入力パッチに対する高解像パッチｙＨを生成する。図５に線形和で生成された高解像パッチ例を示す。
ここで、入力パッチの平均値に合わせ、明るさを補正する。この処理をすべての入力パッチに適用し、高解像パッチが重複する領域はその平均値を求めることにより、高解像画像を生成する。 Here, the brightness is corrected according to the average value of the input patches. This process is applied to all input patches, and a high-resolution image is generated by obtaining an average value of regions where the high-resolution patches overlap.
高画質に拡大できることで知られ、医用画像の拡大にも使用される双三次内挿法と2次元高解像技術を比較することにより、高解像化における高解像技術の有用性を評価した。実験試料は、三重大学医学部附属病院の心臓用３２チャンネルコイルを備えたＭＲ撮影装置で、自由呼吸下、ナビゲーターエコー法による呼吸同期、心電図同期、Ｔ２prep法による静脈血や心筋信号抑制、ＳＰＩＲ法による脂肪信号の抑制の条件の下、撮影された３５患者（１．５Ｔ：１６患者、３．０Ｔ：１９患者）の冠動脈ＭＲＡ(Magnetic
Resonance Angiography)である。各冠動脈ＭＲＡは、５１２×５１２のスライス画像サイズで、１５０枚のスライス画像で構成される。
Evaluates the usefulness of high-resolution technology in high-resolution by comparing bicubic interpolation, which is well-known for high-quality enlargement and is also used for medical image enlargement, and two-dimensional high-resolution technology did. The experimental sample is an MR imaging device equipped with a 32-channel coil for the heart of the Mie University Medical Hospital. Respiratory synchronization using the navigator echo method, electrocardiogram synchronization, venous blood and myocardial signal suppression using the T2prep method, and SPIR method Coronary artery MRA (Magnetic) of 35 patients (1.5T: 16 patients, 3.0T: 19 patients) taken under conditions of fat signal suppression
Resonance Angiography). Each coronary artery MRA has a slice image size of 512 × 512 and includes 150 slice images.
まず、２次元高解像技術による高解像化の正確さを評価するため、５１２×５１２サイズのＭＲＡ画像をダウンサンプリングすることにより、２５６×２５６サイズの縮小画像を生成した。次に、この縮小画像を２次元高解像技術と双三次内挿法を用いて５１２×５１２の画像サイズに復元し、元画像と各復元画像のＲＭＳＥ(Root
Mean Squared Error)、ＰＳＮＲ(Peak
Signal-to-Noise Ratio)、ＳＳＩＭ(Structural
Similarity Index)により、その忠実性を比較した。２次元高解像技術では、各スライス画像を入力画像（２５６×２５６）とし、入力画像を中心とした上下２スライスの計５スライス画像を学習用高解像画像（２５６×２５６×５）として用い、それらをaxial面で１／２のサイズにダウンサンプリングすることにより学習用低解像画像（１２８×１２８×５）を生成した。
First, a 256 × 256 size reduced image was generated by down-sampling a 512 × 512 size MRA image in order to evaluate the accuracy of high resolution by the two-dimensional high resolution technology. Next, this reduced image is restored to an image size of 512 × 512 using a two-dimensional high resolution technique and bicubic interpolation, and the RMSE (Root) of the original image and each restored image is restored.
Mean Squared Error), PSNR (Peak
Signal-to-Noise Ratio), SSIM (Structural
Similarity Index) compared their fidelity. In the two-dimensional high-resolution technique, each slice image is set as an input image (256 × 256), and a total of five slice images of two upper and lower slices centering on the input image are used as a learning high-resolution image (256 × 256 × 5). The low resolution images for learning (128 × 128 × 5) were generated by downsampling them to 1/2 the size on the axial plane.
そして、学習用低解像画像から切り出した５×５サイズの学習用低解像パッチと、それに対応する位置にある学習用高解像画像の１０×１０サイズの学習用高解像パッチをペアにし、それを辞書に格納した。次に、入力画像を各画素で５×５サイズのパッチに切り出し、各入力パッチに対して辞書から特徴量空間のユークリッドの距離が最小となる学習用低解像パッチを類似パッチとして７個選択した。ここで、特徴量空間は５×５のパッチサイズにおける移動平均値とｘ方向、ｙ方向の各１次差分値および各２次差分値により５×５×５次元の特徴量空間で構成された。そして、類似パッチの線形和で入力パッチを復元するための重み係数を最小二乗法で計算し、その重み係数と類似パッチに対応する学習用高解像パッチの線形和で、入力パッチに対する高解像パッチを生成した。 Then, a 5 × 5 size learning low resolution patch cut out from the learning low resolution image is paired with a 10 × 10 size learning high resolution patch of the learning high resolution image at the corresponding position. And stored it in the dictionary. Next, the input image is cut into 5 × 5 size patches for each pixel, and seven low-resolution patches for learning that minimize the Euclidean distance from the dictionary to the feature amount space are selected as similar patches for each input patch. did. Here, the feature amount space is composed of a feature amount space of 5 × 5 × 5 dimensions by the moving average value in the 5 × 5 patch size, each primary difference value in the x direction and y direction, and each secondary difference value. . Then, a weighting factor for restoring the input patch by the linear sum of the similar patches is calculated by the least square method, and the high-resolution for the input patch is calculated by the linear sum of the weighting factor and the learning high-resolution patch corresponding to the similar patch. An image patch was generated.
この処理をすべての入力パッチに適用し、高解像パッチが重複する領域はその平均値を求めることにより、２倍の高解像画像を生成した。その結果、２次元高解像技術のＲＭＳＥ、ＰＳＮＲ、ＳＳＩＭの平均は２．５２、２２．２２ｄＢ、０．９８９、双三次内挿法は３．０８、２０．２ｄＢ、０．９８４であり、各指標において２次元高解像技術が双三次内挿法より復元の忠実性が有意に向上した（Ｐ < ０．００１）。また、図６のように高解像技術による画像は、境界のボケを抑制しながらノイズを減少した高画質な画像であった。 This process was applied to all the input patches, and the area where the high resolution patches overlap was obtained by calculating the average value, thereby generating a double high resolution image. As a result, the average of RMSE, PSNR, and SSIM of the two-dimensional high resolution technique is 2.52, 22.22 dB, 0.989, and the bicubic interpolation method is 3.08, 22.2 dB, 0.984, In each index, the fidelity of restoration of the two-dimensional high-resolution technique is significantly improved over the bicubic interpolation method (P <0.001). Further, as shown in FIG. 6, the image by the high resolution technique is a high-quality image in which noise is reduced while suppressing blurring at the boundary.
次に、２次元高解像技術と双三次内挿法により冠動脈ＭＲＡ（５１２×５１２）をそれぞれ１０２４×１０２４の画像サイズへ高解像化（拡大）し、冠動脈狭窄検出における診断能を観察者実験により比較した。ここで、入力パッチと学習用低解像パッチのサイズを7×7、学習用高解像パッチのサイズを１４×１４で与えた以外は、上記２次元高解像技術と同様である。観察者実験は高解像技術による画像と双三次内挿法による画像を用いた２セッションからなる。実験試料は上記３５患者の冠動脈ＭＲＡを用い、各冠動脈ＭＲＡは観察者実験用のグラフィカルユーザインタフェースにおいてaxial画像として表示され、自由なページングを可能とした。 Next, the two-dimensional high-resolution technique and bicubic interpolation are used to increase (enlarge) the coronary artery MRA (512 x 512) to an image size of 1024 x 1024, and the diagnostic ability in detecting coronary artery stenosis Comparison was made by experiment. Here, the two-dimensional high-resolution technique is the same as that described above except that the size of the input patch and the learning low-resolution patch is 7 × 7 and the size of the learning high-resolution patch is 14 × 14. The observer experiment consists of two sessions using an image by high resolution technology and an image by bicubic interpolation. The coronary artery MRA of the 35 patients was used as an experimental sample, and each coronary artery MRA was displayed as an axial image in a graphical user interface for an observer experiment, enabling free paging.
各症例において、３人の読影者が独立して、５０％以上狭窄の有無に対する確信度を０．０（絶対にない）〜１．０（絶対にある）の連続スケールで評価した。ＲＯＣ(Receiver
Operating Characteristic) 解析の結果を図７に示す。ＲＯＣ曲線下の面積の平均は、高解像技術で０．８６１、双三次内挿法で０．７９７が得られ、高解像技術における診断能が有意に優れていた（Ｐ=０．０２４)。また、高解像技術の使用により観察者間変動は０．１７０から０．１６４に、級内相関係数は０．８１２から０．８５５へ改善された。
In each case, three radiographers independently evaluated the confidence for the presence or absence of stenosis by 50% or more on a continuous scale of 0.0 (absolutely not) to 1.0 (absolutely). ROC (Receiver
The results of the Operating Characteristic analysis are shown in FIG. The average area under the ROC curve was 0.861 with the high resolution technique and 0.797 with the bicubic interpolation method, and the diagnostic ability in the high resolution technique was significantly superior (P = 0.024). ). In addition, the use of a high resolution technique improved the inter-observer variation from 0.170 to 0.164, and the intraclass correlation coefficient from 0.812 to 0.855.
双三次内挿法と３次元高解像技術を比較することにより、高解像化における高解像技術の有用性を評価した。実験試料は、三重大学医学部附属病院の心臓用３２チャンネルコイルを備えた１．５Ｔ ＭＲ撮影装置で、自由呼吸下、ナビゲーターエコー法による呼吸同期、心電図同期、Ｔ２ prep法による静脈血や心筋信号抑制、ＳＰＩＲ法による脂肪信号の抑制の条件の下、撮影された４６患者の冠動脈ＭＲＡである。各冠動脈ＭＲＡは、５１２×５１２のスライス画像サイズで、１５０枚のスライス画像で構成される。 By comparing the bicubic interpolation method and the three-dimensional high-resolution technique, the usefulness of the high-resolution technique in high resolution was evaluated. The experimental sample is a 1.5T MR imaging device equipped with a 32-channel coil for the heart of the Mie University Medical Hospital. Respiratory synchronization using the navigator echo method, electrocardiogram synchronization, and venous blood and myocardial signal suppression using the T2 prep method. , 46 patients coronary artery MRA imaged under the condition of suppression of fat signal by SPIR method. Each coronary artery MRA has a slice image size of 512 × 512 and is composed of 150 slice images.
まず、３次元高解像技術による高解像化の正確さを評価するため、５１２×５１２×１５０サイズのＭＲＡ画像をダウンサンプリングすることにより、２５６×２５６×７５サイズの縮小画像を生成した。次に、この縮小画像を３次元高解像技術と双三次内挿法を用いて５１２×５１２×１５０の画像サイズに復元し、元画像と各復元画像のＲＭＳＥ、ＰＳＮＲ、ＳＳＩＭにより、その忠実性を比較した。３次元高解像技術では、各スライス画像を中心としたボリューム画像（２５６×２５６×５）を入力ボリューム画像とし、入力ボリューム画像を中心に１４スライス画像を学習用高解像画像（２５６×２５６×１４）として用い、それらを１／２のサイズにダウンサンプリングすることにより学習用低解像画像（１２８×１２８×７）を生成した。 First, a 256 × 256 × 75 size reduced image was generated by down-sampling a 512 × 512 × 150 size MRA image in order to evaluate the accuracy of high resolution by the three-dimensional high resolution technology. Next, the reduced image is restored to an image size of 512 × 512 × 150 using a three-dimensional high resolution technique and bicubic interpolation, and the original image and each restored image are faithfully reproduced by RMSE, PSNR, and SSIM. Sex was compared. In the three-dimensional high resolution technology, a volume image (256 × 256 × 5) centered on each slice image is used as an input volume image, and a 14 slice image centered on the input volume image is used as a learning high resolution image (256 × 256). × 14), and a low resolution image for learning (128 × 128 × 7) was generated by down-sampling them to 1/2 size.
そして、学習用低解像画像から切り出した５×５×５サイズの学習用低解像パッチと、それに対応する位置にある学習用高解像画像の１０×１０×１０サイズの学習用高解像パッチをペアにし、それを辞書に格納した。次に、入力ボリューム画像を各画素で５×５×５サイズのパッチに切り出し、各入力パッチに対して辞書から特徴量空間のユークリッドの距離が最小となる学習用低解像パッチを類似パッチとして７個選択した。ここで、特徴量空間は５×５×５のパッチサイズにおける移動平均値とｘ方向、ｙ方向、ｚ方向の各２次差分値により５×５×５×４次元の特徴量空間で構成された。そして、類似パッチの線形和で入力パッチを復元するための重み係数を最小二乗法で計算し、その重み係数と類似パッチに対応する学習用高解像パッチの線形和で、入力パッチに対する高解像パッチを生成した。 Then, a learning low resolution patch of 5 × 5 × 5 size cut out from the learning low resolution image and a learning high resolution image of 10 × 10 × 10 size of the learning high resolution image at the corresponding position The image patches were paired and stored in the dictionary. Next, the input volume image is cut into 5 × 5 × 5 size patches for each pixel, and a learning low resolution patch that minimizes the distance of the Euclidean feature space from the dictionary to each input patch is used as a similar patch. Seven were selected. Here, the feature amount space is composed of a 5 × 5 × 5 × 4 dimensional feature amount space by a moving average value in a patch size of 5 × 5 × 5 and secondary difference values in the x, y, and z directions. It was. Then, a weighting factor for restoring the input patch by the linear sum of the similar patches is calculated by the least square method, and the high-resolution for the input patch is calculated by the linear sum of the weighting factor and the learning high-resolution patch corresponding to the similar patch. An image patch was generated.
この処理をすべての入力パッチに適用し、高解像パッチが重複する領域はその平均値を求めることにより、２倍の高解像画像を生成した。その結果、３次元高解像技術のＲＭＳＥ、ＰＳＮＲ、ＳＳＩＭの平均は２．７５、２３．６ｄＢ、０．９８７、双三次内挿法は３．５７、２０．８ｄＢ、０．９８１であり、各指標において３次元高解像技術が双三次内挿法より復元の忠実性が有意に向上した（Ｐ＜０．００１）。また、実施例１の２次元高解像技術を実験試料に適用した場合のＲＭＳＥ、ＰＳＮＲ、ＳＳＩＭの平均は３．２８、２１．４ｄＢ、０．９８２（Ｐ＜０．０００１）であり、２次元高解像技術に比べ、３次元高解像技術が冠動脈ＭＲＡの高解像化に、より有用となる可能性が示唆された。 This process was applied to all the input patches, and the area where the high resolution patches overlap was obtained by calculating the average value, thereby generating a double high resolution image. As a result, the average of RMSE, PSNR, and SSIM of the three-dimensional high resolution technology is 2.75, 23.6 dB, 0.987, and the bicubic interpolation method is 3.57, 20.8 dB, 0.981, In each index, the fidelity of restoration of the three-dimensional high-resolution technique is significantly improved over the bicubic interpolation method (P <0.001). In addition, the average of RMSE, PSNR, and SSIM when the two-dimensional high-resolution technique of Example 1 is applied to an experimental sample is 3.28, 21.4 dB, 0.982 (P <0.0001), and 2 Compared to the three-dimensional high-resolution technique, it is suggested that the three-dimensional high-resolution technique may be more useful for increasing the resolution of the coronary artery MRA.
次に、３次元高解像技術と双三次内挿法により冠動脈ＭＲＡ（５１２×５１２×１５０）をそれぞれ１０２４×１０２４×３００の画像サイズへ高解像化（拡大）し、冠動脈狭窄検出における診断能を観察者実験により比較した。ここで、入力パッチと学習用低解像パッチのサイズを７×７×７、学習用高解像パッチのサイズを１４×１４×１４で与えた以外は、上記３次元高解像技術と同様である。 Next, the three-dimensional high-resolution technique and bicubic interpolation are used to increase (enlarge) the coronary artery MRA (512 × 512 × 150) to an image size of 1024 × 1024 × 300, respectively, for diagnosis in detecting coronary artery stenosis. The ability was compared by an observer experiment. Here, except that the size of the input patch and the learning low resolution patch is 7 × 7 × 7, and the size of the learning high resolution patch is 14 × 14 × 14, the same as the above three-dimensional high resolution technology. It is.
観察者実験は３次元高解像技術による画像と双三次内挿法による画像を用いた２セッションからなる。実験試料は上記４６患者の冠動脈ＭＲＡを用い、各冠動脈ＭＲＡは商用ＤＩＣＯＭ(Digital-Imaging-and-Communication-in-Medicine)Viewer上で、sliding ＳＬＡＢ ＭＩＰ法により読影された。各症例において、３人の読影者が独立して、５０％以上狭窄の有無に対する確信度を０．０（絶対にない）〜１．０（絶対にある）の連続スケールで評価した。ＲＯＣ解析の結果を図８に示す。ＲＯＣ曲線下の面積の平均は、高解像技術で０．８４０、双三次内挿法で０．７９２が得られ、高解像技術における診断能が有意に優れていた（Ｐ=０．０２０)。また、高解像技術の使用により級内相関係数は０．５７９から０．６７０へ改善された。
＜変形例＞
The observer experiment consists of two sessions using an image by a three-dimensional high-resolution technique and an image by a bicubic interpolation method. The coronary artery MRA of the above 46 patients was used as an experimental sample, and each coronary artery MRA was read by a sliding SLAB MIP method on a commercial DICOM (Digital-Imaging-and-Communication-in-Medicine) Viewer. In each case, three radiographers independently evaluated the confidence for the presence or absence of stenosis by 50% or more on a continuous scale of 0.0 (absolutely not) to 1.0 (absolutely). The result of the ROC analysis is shown in FIG. The average area under the ROC curve was 0.840 with the high resolution technique and 0.792 with the bicubic interpolation method, and the diagnostic ability in the high resolution technique was significantly superior (P = 0.020). ). In addition, the intraclass correlation coefficient was improved from 0.579 to 0.670 by using a high resolution technique.
<Modification>
医用画像では高解像化と同様高画質化が強く求められる。
医用画像の高画質化処理方法には、参照する低画質画像と、高画質画像から切り出した学習用パッチのペアで辞書を構築する辞書作成アルゴリズムと、辞書から入力低画質パッチに適した学習用パッチを選択し、それらを合成することにより入力パッチに対する高画質パッチを推定する高解像画像形成方法がある。
図９に高画質化と、高解像化を併記したフローチャートを示す。
また、図１０に辞書作成のための医用高画質・高解像画像形成複合システム５０を示す。医用高画質・高解像画像形成複合システム５０では、次の手順で高画質処理と高解像処理が進められる。
For medical images, high image quality is strongly demanded, as is high resolution.
The medical image quality enhancement processing method includes a dictionary creation algorithm that builds a dictionary with a pair of low-quality images to be referenced and learning patches cut out from the high-quality images, and a learning algorithm suitable for input low-quality patches from the dictionary. There is a high-resolution image forming method for estimating a high-quality patch for an input patch by selecting patches and combining them.
FIG. 9 shows a flowchart in which both high image quality and high resolution are shown.
FIG. 10 shows a combined medical high-quality / high-resolution image forming system 50 for creating a dictionary. In the medical high image quality / high resolution image forming composite system 50, high image quality processing and high resolution processing are performed in the following procedure.
（１）高画質画像５１として、画質が良い２０名程度の別患者画像を学習用画像の高画質画像として用い、それをぼかしたり、ノイズを加えたりして、各学習用高画質画像に対し複数種類の学習用低画質画像５２を生成する。ぼかしの強さは、入力画像の撮影条件、撮影時の身体の動きなどによって異なるため、ガウシアン平滑化フィルタの標準偏差を０画素、３画素、５画素、７画素の４種類としている。また、ノイズの強さは、ガウシアンノイズの標準偏差を０画素、１０画素、２０画素の３種類としている。したがって、１枚の学習用高画質画像から１２枚の学習用低画質画像が生成される。ただし、ガウシアン平滑化フィルタ、ガウシアンノイズの標準偏差がともに０画素の場合は、高画質画像が維持される。高画質画像５１に、ぼかし処理、ノイズ処理を施して低画質画像５２を作成する。 (1) As the high-quality image 51, about 20 different patient images with good image quality are used as the high-quality image of the learning image, and are blurred or added to each learning high-quality image. A plurality of types of learning low-quality images 52 are generated. Since the intensity of the blur varies depending on the shooting conditions of the input image, the body movement at the time of shooting, and the like, the standard deviation of the Gaussian smoothing filter is set to four types of 0 pixel, 3 pixels, 5 pixels, and 7 pixels. Further, the noise intensity has three types of standard deviation of Gaussian noise: 0 pixel, 10 pixel, and 20 pixel. Therefore, twelve low-quality images for learning are generated from one high-quality image for learning. However, when both the Gaussian smoothing filter and the standard deviation of the Gaussian noise are 0 pixels, a high-quality image is maintained. The low-quality image 52 is created by performing blurring processing and noise processing on the high-quality image 51.
（２）低画質画像５２をN×N(×N)サイズの低画質画像パッチ５４に分割する。
（３）低画質画像パッチ５４の位置に対応させてN×N(×N)サイズの高画質画像パッチ５５を作成する。
（４）低画質画像パッチ５４と高画質画像パッチ５５とを対応付けさせた辞書５３を作成し、記憶部に記憶させる。
（５）入力画像５６をパッチに分割して入力画像パッチ５７を作成する。各入力パッチと辞書中の低画質画像パッチ５４のすべてと個別に対比させて類似性評価を行い、類似性評価値の高いパッチ順に所定個数の低画質パッチを抽出する。
(2) The low-quality image 52 is divided into N × N (× N) size low-quality image patches 54.
(3) An N × N (× N) size high-quality image patch 55 is created in correspondence with the position of the low-quality image patch 54.
(4) A dictionary 53 in which the low-quality image patch 54 and the high-quality image patch 55 are associated is created and stored in the storage unit.
(5) The input image 56 is divided into patches to create an input image patch 57. Similarity evaluation is performed by individually comparing each input patch and all the low-quality image patches 54 in the dictionary, and a predetermined number of low-quality patches are extracted in the order of patches having the highest similarity evaluation value.
（６）抽出した低画質画像パッチ５４に対応する高画質画像パッチ５５を抽出する。
（７）抽出した高画質画像パッチ５５に重みづけした線形和により画像を合成し、合成した画像を入力画像パッチに置き換える。
（８）この置き換え処理を全ての入力画像パッチについて実施して、入力画像５６に対応する高画質画像５８を得る。
（９）高画質画像５８を図１、２に示す医用高解像画像形成システムに入力して、高画質化処理を施す。
(6) The high-quality image patch 55 corresponding to the extracted low-quality image patch 54 is extracted.
(7) An image is synthesized by a linear sum weighted to the extracted high-quality image patch 55, and the synthesized image is replaced with an input image patch.
(8) This replacement process is performed on all input image patches to obtain a high-quality image 58 corresponding to the input image 56.
(9) The high-quality image 58 is input to the medical high-resolution image forming system shown in FIGS.
なお、図１０では、高画質化処理の後に高解像化処理を実施しているが、順序を逆にして高解像化処理の後に高画質化処理してもよい。 In FIG. 10, the high-resolution processing is performed after the high-quality processing, but the order may be reversed and the high-quality processing may be performed after the high-resolution processing.
本発明に関わる医用高解像画像形成システムを用いることにより、画質や解像度の向上が望まれるＣＴ画像、ＭＲ画像、超音波画像をはじめとするあらゆる医用画像に対し、後処理としてソフトウェア的に高画質化、高解像化することが可能である。したがって、独立した医用画像処理ソフトとしてだけではなく、撮影機器装置、医療用画像管理システム（ＰＡＣＳ:Picture
Archiving and Communication System）、ＤＩＣＯＭ Viewer、医療用高精細モニタに、高画質化、高解像化機能として導入することができる。
By using the medical high-resolution image forming system according to the present invention, all medical images including CT images, MR images, and ultrasonic images, for which improvement in image quality and resolution is desired, are highly processed as software. It is possible to improve image quality and resolution. Therefore, not only as an independent medical image processing software, but also as an imaging device, a medical image management system (PACS: Picture
Archiving and Communication System), DICOM Viewer, and medical high-definition monitors can be introduced as high-quality and high-resolution functions.
１０ 医用高解像画像形成システム
１１ 学習用高解像画像入力部
１２ 辞書作成部
１３ 辞書記憶部
１４ 画像入力部
１５ 高解像出力画像作成部
１６ 学習用高解像画像
１７ 入力画像
１８ 高解像出力画像
２０ 高解像画像形成システム
２１ 学習用低解像画像
２２ 辞書
２３ 学習用低解像パッチ
２４ 学習用高解像パッチ
２５ 入力画像パッチ
５０ 医用高画質・高解像画像形成複合システム
DESCRIPTION OF SYMBOLS 10 Medical high resolution image formation system 11 Learning high resolution image input part 12 Dictionary creation part 13 Dictionary memory | storage part 14 Image input part 15 High resolution output image creation part 16 High resolution image for learning 17 Input image 18 High resolution Image output image 20 High resolution image forming system 21 Low resolution image for learning 22 Dictionary 23 Low resolution patch for learning 24 High resolution patch for learning 25 Input image patch 50 High image quality / high resolution image forming combined system
Claims (4)
１つ以上の学習用高解像画像の画素数をＭ分の１に縮小した学習用低解像画像を作成し、
該学習用低解像画像をＮ×Ｎサイズの学習用低解像画像パッチに分割し、該学習用低解像画像パッチに対応させて前記学習用高解像画像をＭＮ×ＭＮサイズの学習用高解像画像パッチに分割し、
前記学習用低解像画像パッチと前記学習用高解像画像パッチとを対応付けさせて作成される辞書を作成する辞書作成部と、
該辞書作成部により作成される辞書を記憶する辞書記憶部と、
低解像の入力画像を入力する画像入力部と、
入力画像をＮ×Ｎサイズの入力画像パッチに分割し、分割された各入力画像パッチと、前記辞書の前記学習用低解像画像パッチの全てとを個別に対比させて類似性評価を行い、類似性評価値の高い順にＫ個の学習用低解像パッチを抽出し、該抽出されたＫ個の前記学習用低解像パッチに対応づけされた前記学習用高解像パッチを抽出し、該抽出されたＫ個の前記学習用高解像パッチを結合して高解像出力画像を作成し出力する高解像出力画像作成部と、
を有することを特徴とする医用高解像画像形成システム。 Medical high-resolution image formation that uses a low-resolution medical image created by a computed tomography (CT) system and a magnetic resonance imaging (MRI) system as an input image, and outputs the input image as a high-resolution output image A system,
Creating a learning low-resolution image in which the number of pixels of one or more learning high-resolution images is reduced to 1 / M,
The learning low-resolution image is divided into N × N-size learning low-resolution image patches, and the learning high-resolution image is matched with the learning low-resolution image patches so that the learning high-resolution image is learned in MN × MN size. Split into high resolution image patches for
A dictionary creation unit that creates a dictionary created by associating the learning low-resolution image patch and the learning high-resolution image patch;
A dictionary storage unit for storing a dictionary created by the dictionary creation unit;
An image input unit for inputting a low-resolution input image;
The input image is divided into N × N size input image patches, and each divided input image patch is individually compared with all of the learning low-resolution image patches in the dictionary to perform similarity evaluation, Extracting K learning low-resolution patches in descending order of similarity evaluation value, and extracting the learning high-resolution patches associated with the extracted K low-resolution patches for learning; A high-resolution output image creation unit that creates and outputs a high-resolution output image by combining the extracted K high-resolution patches for learning; and
A medical high-resolution image forming system.
１つ以上の学習用高解像画像の画素数をＭ分の１に縮小した学習用低解像画像を作成し、
該学習用低解像画像をＮ×Ｎサイズの学習用低解像画像パッチに分割し、該学習用低解像画像パッチに対応させて前記学習用高解像画像をＭＮ×ＭＮサイズの学習用高解像画像パッチに分割し、
前記学習用低解像画像パッチと前記学習用高解像画像パッチとを対応付けさせて作成される辞書を作成して、該辞書を辞書記憶部に記憶し、
低解像の入力画像を入力して、該入力画像をＮ×Ｎサイズの入力画像パッチに分割し、分割された各入力画像パッチと、前記辞書の前記学習用低解像画像パッチの全てとを個別に対比させて類似性評価を行い、類似性評価値の高い順にＫ個の学習用低解像画像パッチを抽出し、
該抽出されたＫ個の前記学習用低解像パッチに対応づけされた前記学習用高解像パッチを抽出し、該抽出されたＫ個の前記学習用高解像パッチを結合して高解像出力画像を作成し出力することを特徴とする医用高解像画像形成方法。 Medical high-resolution image formation that uses a low-resolution medical image created by a computed tomography (CT) system and a magnetic resonance imaging (MRI) system as an input image, and outputs the input image as a high-resolution output image A method,
Creating a learning low-resolution image in which the number of pixels of one or more learning high-resolution images is reduced to 1 / M,
The learning low-resolution image is divided into N × N-size learning low-resolution image patches, and the learning high-resolution image is matched with the learning low-resolution image patches so that the learning high-resolution image is learned in MN × MN size. Split into high resolution image patches for
Creating a dictionary created by associating the learning low-resolution image patch and the learning high-resolution image patch, and storing the dictionary in a dictionary storage unit;
A low resolution input image is input, the input image is divided into N × N size input image patches, each divided input image patch, and all of the learning low resolution image patches of the dictionary, Are compared individually, K low-resolution image patches for learning are extracted in descending order of similarity evaluation value,
The learning high-resolution patches associated with the extracted K low-resolution patches for learning are extracted, and the K high-resolution patches for learning are combined to obtain a high resolution. A medical high-resolution image forming method characterized in that an image output image is created and output.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| JP2014000148A JP2015129987A (en) | 2014-01-06 | 2014-01-06 | System and method of forming medical high-resolution image |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| JP2014000148A JP2015129987A (en) | 2014-01-06 | 2014-01-06 | System and method of forming medical high-resolution image |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| JP2015129987A true JP2015129987A (en) | 2015-07-16 |
Family
ID=53760687
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| JP2014000148A Pending JP2015129987A (en) | 2014-01-06 | 2014-01-06 | System and method of forming medical high-resolution image |
Country Status (1)
| Country | Link |
|---|---|
| JP (1) | JP2015129987A (en) |
Cited By (38)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2017050222A (en) * | 2015-09-03 | 2017-03-09 | 株式会社東芝 | Inspection apparatus and inspection method |
| JP2017091529A (en) * | 2015-11-09 | 2017-05-25 | トムソン ライセンシングＴｈｏｍｓｏｎ Ｌｉｃｅｎｓｉｎｇ | Method for upscaling noisy images, and apparatus for upscaling noisy images |
| JP2017113182A (en) * | 2015-12-22 | 2017-06-29 | 高知県公立大学法人 | Super-resolution system of mr image |
| JP2017185007A (en) * | 2016-04-05 | 2017-10-12 | 株式会社島津製作所 | Radiographic apparatus, radiation image object detection program, and object detection method in radiation image |
| FR3050826A1 (en) * | 2016-05-02 | 2017-11-03 | Safran | METHOD OF NON-DESTRUCTIVE CONTROL BY LEARNING |
| JP6236731B1 (en) * | 2017-03-10 | 2017-11-29 | 株式会社ツバサファクトリー | Super-resolution processing apparatus, super-resolution processing method, and computer program |
| KR20190001608A (en) * | 2016-05-25 | 2019-01-04 | 케이엘에이-텐코 코포레이션 | Generating simulated images from input images for semiconductor applications |
| WO2019026407A1 (en) * | 2017-07-31 | 2019-02-07 | 株式会社日立製作所 | Medical imaging device and medical image processing method |
| KR20190045884A (en) * | 2017-10-24 | 2019-05-03 | 울산대학교 산학협력단 | Apparatus for interpreting medical image and method for providing disease information |
| CN109816592A (en) * | 2019-01-26 | 2019-05-28 | 福州大学 | Single-frame image continuous scale super-resolution method based on convolutional neural network |
| KR20190114229A (en) * | 2018-03-29 | 2019-10-10 | 울산대학교 산학협력단 | Apparatus and method for processing image |
| KR20190137283A (en) * | 2018-06-01 | 2019-12-11 | 연세대학교 산학협력단 | Method for producing medical image and device for producing medical image |
| JP2019212050A (en) * | 2018-06-05 | 2019-12-12 | 株式会社島津製作所 | Image processing method, image processing device and learning model creation method |
| WO2019240257A1 (en) * | 2018-06-15 | 2019-12-19 | キヤノン株式会社 | Medical image processing device, medical image processing method and program |
| JP2019216848A (en) * | 2018-06-15 | 2019-12-26 | キヤノン株式会社 | Medical image processing apparatus, medical image processing method, and program |
| JP2020010823A (en) * | 2018-07-18 | 2020-01-23 | キヤノンメディカルシステムズ株式会社 | Medical information processing apparatus, medical information processing system, and medical information processing program |
| US10559074B2 (en) | 2017-02-20 | 2020-02-11 | Hitachi High-Technologies Corporation | Sample observation device and sample observation method |
| JP2020025786A (en) * | 2018-08-14 | 2020-02-20 | 富士フイルム株式会社 | Image processing apparatus, method and program |
| JP2020046404A (en) * | 2018-09-21 | 2020-03-26 | キヤノンメディカルシステムズ株式会社 | Medical image processor |
| JP2020058800A (en) * | 2018-10-10 | 2020-04-16 | キヤノン株式会社 | Image processing device, image processing method, and image processing program |
| JP2020064609A (en) * | 2018-08-23 | 2020-04-23 | ゼネラル・エレクトリック・カンパニイ | Patient-specific deep learning image denoising methods and systems |
| JP2020516345A (en) * | 2017-04-05 | 2020-06-11 | ゼネラル・エレクトリック・カンパニイ | Tomography reconstruction based on deep learning |
| JP2020114295A (en) * | 2019-01-17 | 2020-07-30 | キヤノンメディカルシステムズ株式会社 | Ultrasonic diagnostic apparatus and learning program |
| JP2020114294A (en) * | 2019-01-17 | 2020-07-30 | キヤノンメディカルシステムズ株式会社 | Ultrasonic diagnostic apparatus and learning program |
| JP2020144700A (en) * | 2019-03-07 | 2020-09-10 | 株式会社日立製作所 | Diagnostic imaging apparatus, image processing method and program |
| JP2020162802A (en) * | 2019-03-29 | 2020-10-08 | ゼネラル・エレクトリック・カンパニイ | Ultrasonic device and control program thereof |
| JP2020166813A (en) * | 2019-03-11 | 2020-10-08 | キヤノン株式会社 | Medical image processing device, medical image processing method, and program |
| JP2021007017A (en) * | 2020-09-15 | 2021-01-21 | 株式会社トプコン | Medical image processing method and medical image processing device |
| JPWO2021033291A1 (en) * | 2019-08-21 | 2021-02-25 | ||
| JP2021115225A (en) * | 2020-01-24 | 2021-08-10 | キヤノン株式会社 | Ultrasonic diagnostic device, learning device, image processing method, and program |
| JP2021133142A (en) * | 2020-02-28 | 2021-09-13 | 株式会社日立製作所 | Medical imaging apparatus, medical image processing apparatus and image processing program |
| US20210398259A1 (en) | 2019-03-11 | 2021-12-23 | Canon Kabushiki Kaisha | Image processing apparatus, image processing method, and storage medium |
| CN116310959A (en) * | 2023-02-21 | 2023-06-23 | 南京智蓝芯联信息科技有限公司 | Method and system for identifying low-quality camera picture in complex scene |
| JP7324195B2 (en) | 2017-10-23 | 2023-08-09 | コーニンクレッカ フィリップス エヌ ヴェ | Optimizing Positron Emission Tomography System Design Using Deep Imaging |
| US11730388B2 (en) | 2019-06-25 | 2023-08-22 | Siemens Healthcare Gmbh | MRI image reconstruction using machine learning |
| CN116805284A (en) * | 2023-08-28 | 2023-09-26 | 之江实验室 | Feature migration-based super-resolution reconstruction method and system between three-dimensional magnetic resonance planes |
| US11922601B2 (en) | 2018-10-10 | 2024-03-05 | Canon Kabushiki Kaisha | Medical image processing apparatus, medical image processing method and computer-readable medium |
| US11961209B2 (en) | 2019-10-20 | 2024-04-16 | Canon Medical Systems Corporation | Noise to noise ensemble learning for pet and CT image and data denoising |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2002325761A (en) * | 2000-06-30 | 2002-11-12 | Hitachi Medical Corp | Image diagnosis supporting device |
| WO2013089263A1 (en) * | 2011-12-12 | 2013-06-20 | 日本電気株式会社 | Image processing device, and image processing method |
-
2014
- 2014-01-06 JP JP2014000148A patent/JP2015129987A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2002325761A (en) * | 2000-06-30 | 2002-11-12 | Hitachi Medical Corp | Image diagnosis supporting device |
| WO2013089263A1 (en) * | 2011-12-12 | 2013-06-20 | 日本電気株式会社 | Image processing device, and image processing method |
Non-Patent Citations (1)
| Title |
|---|
| 谷口 和輝: "多様体学習法と事例ベース法を融合した超解像技術とその高速化", 電気学会論文誌Ｃ, vol. 132, no. 11, JPN6017042379, 1 November 2012 (2012-11-01), pages 1768 - 1773, ISSN: 0003675730 * |
Cited By (72)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US10096100B2 (en) | 2015-09-03 | 2018-10-09 | Toshiba Memory Corporation | Inspection device, inspection method, and image processing program |
| JP2017050222A (en) * | 2015-09-03 | 2017-03-09 | 株式会社東芝 | Inspection apparatus and inspection method |
| JP2017091529A (en) * | 2015-11-09 | 2017-05-25 | トムソン ライセンシングＴｈｏｍｓｏｎ Ｌｉｃｅｎｓｉｎｇ | Method for upscaling noisy images, and apparatus for upscaling noisy images |
| JP7260243B2 (en) | 2015-11-09 | 2023-04-18 | インターデジタル マディソン パテント ホールディングス， エスアーエス | Noisy image upscaling method and apparatus for noisy image upscaling |
| JP2017113182A (en) * | 2015-12-22 | 2017-06-29 | 高知県公立大学法人 | Super-resolution system of mr image |
| JP2017185007A (en) * | 2016-04-05 | 2017-10-12 | 株式会社島津製作所 | Radiographic apparatus, radiation image object detection program, and object detection method in radiation image |
| FR3050826A1 (en) * | 2016-05-02 | 2017-11-03 | Safran | METHOD OF NON-DESTRUCTIVE CONTROL BY LEARNING |
| KR20190001608A (en) * | 2016-05-25 | 2019-01-04 | 케이엘에이-텐코 코포레이션 | Generating simulated images from input images for semiconductor applications |
| KR102229081B1 (en) * | 2016-05-25 | 2021-03-16 | 케이엘에이 코포레이션 | Generate simulated images from input images for semiconductor applications |
| US10559074B2 (en) | 2017-02-20 | 2020-02-11 | Hitachi High-Technologies Corporation | Sample observation device and sample observation method |
| US11170483B2 (en) | 2017-02-20 | 2021-11-09 | Hitachi High-Tech Corporation | Sample observation device and sample observation method |
| JP2018151747A (en) * | 2017-03-10 | 2018-09-27 | 株式会社ツバサファクトリー | Super-resolution processing apparatus, super-resolution processing method and computer program |
| JP6236731B1 (en) * | 2017-03-10 | 2017-11-29 | 株式会社ツバサファクトリー | Super-resolution processing apparatus, super-resolution processing method, and computer program |
| JP7187476B2 (en) | 2017-04-05 | 2022-12-12 | ゼネラル・エレクトリック・カンパニイ | Tomographic reconstruction based on deep learning |
| JP2020516345A (en) * | 2017-04-05 | 2020-06-11 | ゼネラル・エレクトリック・カンパニイ | Tomography reconstruction based on deep learning |
| WO2019026407A1 (en) * | 2017-07-31 | 2019-02-07 | 株式会社日立製作所 | Medical imaging device and medical image processing method |
| US11100638B2 (en) | 2017-07-31 | 2021-08-24 | Hitachi, Ltd. | Medical imaging device and medical image processing method using a neutral network |
| JP7324195B2 (en) | 2017-10-23 | 2023-08-09 | コーニンクレッカ フィリップス エヌ ヴェ | Optimizing Positron Emission Tomography System Design Using Deep Imaging |
| KR102181701B1 (en) * | 2017-10-24 | 2020-11-24 | 울산대학교 산학협력단 | Apparatus for interpreting medical image and method for providing disease information |
| KR20190045884A (en) * | 2017-10-24 | 2019-05-03 | 울산대학교 산학협력단 | Apparatus for interpreting medical image and method for providing disease information |
| KR20190114229A (en) * | 2018-03-29 | 2019-10-10 | 울산대학교 산학협력단 | Apparatus and method for processing image |
| KR102084138B1 (en) * | 2018-03-29 | 2020-03-04 | 울산대학교 산학협력단 | Apparatus and method for processing image |
| KR102060895B1 (en) * | 2018-06-01 | 2020-02-11 | 연세대학교 산학협력단 | Method for producing medical image and device for producing medical image |
| KR20190137283A (en) * | 2018-06-01 | 2019-12-11 | 연세대학교 산학협력단 | Method for producing medical image and device for producing medical image |
| JP2022081653A (en) * | 2018-06-05 | 2022-05-31 | 株式会社島津製作所 | Image processing method and image processing device |
| JP2019212050A (en) * | 2018-06-05 | 2019-12-12 | 株式会社島津製作所 | Image processing method, image processing device and learning model creation method |
| JP7405176B2 (en) | 2018-06-05 | 2023-12-26 | 株式会社島津製作所 | Image processing method and image processing device |
| GB2589250A (en) * | 2018-06-15 | 2021-05-26 | Canon Kk | Medical image processing device, medical image processing method and program |
| GB2589250B (en) * | 2018-06-15 | 2023-03-08 | Canon Kk | Medical image processing apparatus, medical image processing method and program |
| WO2019240257A1 (en) * | 2018-06-15 | 2019-12-19 | キヤノン株式会社 | Medical image processing device, medical image processing method and program |
| JP7114358B2 (en) | 2018-06-15 | 2022-08-08 | キヤノン株式会社 | MEDICAL IMAGE PROCESSING APPARATUS, MEDICAL IMAGE PROCESSING METHOD AND PROGRAM |
| JP2019216848A (en) * | 2018-06-15 | 2019-12-26 | キヤノン株式会社 | Medical image processing apparatus, medical image processing method, and program |
| KR102507711B1 (en) * | 2018-06-15 | 2023-03-10 | 캐논 가부시끼가이샤 | Medical image processing apparatus, medical image processing method, and computer readable medium |
| KR20210018461A (en) * | 2018-06-15 | 2021-02-17 | 캐논 가부시끼가이샤 | Medical image processing apparatus, medical image processing method, and computer-readable medium |
| JP7210175B2 (en) | 2018-07-18 | 2023-01-23 | キヤノンメディカルシステムズ株式会社 | Medical information processing device, medical information processing system and medical information processing program |
| JP2020010823A (en) * | 2018-07-18 | 2020-01-23 | キヤノンメディカルシステムズ株式会社 | Medical information processing apparatus, medical information processing system, and medical information processing program |
| JP2020025786A (en) * | 2018-08-14 | 2020-02-20 | 富士フイルム株式会社 | Image processing apparatus, method and program |
| JP2020064609A (en) * | 2018-08-23 | 2020-04-23 | ゼネラル・エレクトリック・カンパニイ | Patient-specific deep learning image denoising methods and systems |
| JP7053541B2 (en) | 2018-08-23 | 2022-04-12 | ゼネラル・エレクトリック・カンパニイ | Patient-specific deep learning image denoising methods and systems |
| JP2020046404A (en) * | 2018-09-21 | 2020-03-26 | キヤノンメディカルシステムズ株式会社 | Medical image processor |
| JP7291466B2 (en) | 2018-09-21 | 2023-06-15 | キヤノンメディカルシステムズ株式会社 | medical image processor |
| JP2020058800A (en) * | 2018-10-10 | 2020-04-16 | キヤノン株式会社 | Image processing device, image processing method, and image processing program |
| US11935241B2 (en) | 2018-10-10 | 2024-03-19 | Canon Kabushiki Kaisha | Image processing apparatus, image processing method and computer-readable medium for improving image quality |
| US11922601B2 (en) | 2018-10-10 | 2024-03-05 | Canon Kabushiki Kaisha | Medical image processing apparatus, medical image processing method and computer-readable medium |
| JP7250653B2 (en) | 2018-10-10 | 2023-04-03 | キヤノン株式会社 | Image processing device, image processing method and program |
| JP7273519B2 (en) | 2019-01-17 | 2023-05-15 | キヤノンメディカルシステムズ株式会社 | Ultrasound diagnostic equipment and learning program |
| JP2020114294A (en) * | 2019-01-17 | 2020-07-30 | キヤノンメディカルシステムズ株式会社 | Ultrasonic diagnostic apparatus and learning program |
| JP2020114295A (en) * | 2019-01-17 | 2020-07-30 | キヤノンメディカルシステムズ株式会社 | Ultrasonic diagnostic apparatus and learning program |
| JP7302972B2 (en) | 2019-01-17 | 2023-07-04 | キヤノンメディカルシステムズ株式会社 | Ultrasound diagnostic equipment and learning program |
| CN109816592A (en) * | 2019-01-26 | 2019-05-28 | 福州大学 | Single-frame image continuous scale super-resolution method based on convolutional neural network |
| CN109816592B (en) * | 2019-01-26 | 2022-05-13 | 福州大学 | Single-frame image continuous scale super-resolution method based on convolutional neural network |
| JP7218215B2 (en) | 2019-03-07 | 2023-02-06 | 株式会社日立製作所 | Image diagnosis device, image processing method and program |
| US11373301B2 (en) | 2019-03-07 | 2022-06-28 | Hitachi, Ltd. | Image diagnostic device, image processing method, and program for obtaining diagnostic prediction models using deep learning |
| JP2020144700A (en) * | 2019-03-07 | 2020-09-10 | 株式会社日立製作所 | Diagnostic imaging apparatus, image processing method and program |
| US11887288B2 (en) | 2019-03-11 | 2024-01-30 | Canon Kabushiki Kaisha | Image processing apparatus, image processing method, and storage medium |
| US20210398259A1 (en) | 2019-03-11 | 2021-12-23 | Canon Kabushiki Kaisha | Image processing apparatus, image processing method, and storage medium |
| JP2020166813A (en) * | 2019-03-11 | 2020-10-08 | キヤノン株式会社 | Medical image processing device, medical image processing method, and program |
| JP7297628B2 (en) | 2019-03-11 | 2023-06-26 | キヤノン株式会社 | MEDICAL IMAGE PROCESSING APPARATUS, MEDICAL IMAGE PROCESSING METHOD AND PROGRAM |
| JP2020162802A (en) * | 2019-03-29 | 2020-10-08 | ゼネラル・エレクトリック・カンパニイ | Ultrasonic device and control program thereof |
| US11730388B2 (en) | 2019-06-25 | 2023-08-22 | Siemens Healthcare Gmbh | MRI image reconstruction using machine learning |
| WO2021033291A1 (en) * | 2019-08-21 | 2021-02-25 | 株式会社島津製作所 | X-ray image capturing method and x-ray image capturing system |
| JPWO2021033291A1 (en) * | 2019-08-21 | 2021-02-25 | ||
| US11961209B2 (en) | 2019-10-20 | 2024-04-16 | Canon Medical Systems Corporation | Noise to noise ensemble learning for pet and CT image and data denoising |
| JP7346314B2 (en) | 2020-01-24 | 2023-09-19 | キヤノン株式会社 | Ultrasonic diagnostic equipment, learning equipment, image processing methods and programs |
| JP2021115225A (en) * | 2020-01-24 | 2021-08-10 | キヤノン株式会社 | Ultrasonic diagnostic device, learning device, image processing method, and program |
| JP7256765B2 (en) | 2020-02-28 | 2023-04-12 | 株式会社日立製作所 | Medical imaging device, medical image processing device, and image processing program |
| JP2021133142A (en) * | 2020-02-28 | 2021-09-13 | 株式会社日立製作所 | Medical imaging apparatus, medical image processing apparatus and image processing program |
| JP2021007017A (en) * | 2020-09-15 | 2021-01-21 | 株式会社トプコン | Medical image processing method and medical image processing device |
| CN116310959A (en) * | 2023-02-21 | 2023-06-23 | 南京智蓝芯联信息科技有限公司 | Method and system for identifying low-quality camera picture in complex scene |
| CN116310959B (en) * | 2023-02-21 | 2023-12-08 | 南京智蓝芯联信息科技有限公司 | Method and system for identifying low-quality camera picture in complex scene |
| CN116805284A (en) * | 2023-08-28 | 2023-09-26 | 之江实验室 | Feature migration-based super-resolution reconstruction method and system between three-dimensional magnetic resonance planes |
| CN116805284B (en) * | 2023-08-28 | 2023-12-19 | 之江实验室 | Feature migration-based super-resolution reconstruction method and system between three-dimensional magnetic resonance planes |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| JP2015129987A (en) | System and method of forming medical high-resolution image | |
| Isaac et al. | Super resolution techniques for medical image processing | |
| Lyu et al. | Super-resolution MRI through deep learning | |
| Xiang et al. | Ultra-fast t2-weighted mr reconstruction using complementary t1-weighted information | |
| McDonagh et al. | Context-sensitive super-resolution for fast fetal magnetic resonance imaging | |
| CA3092994A1 (en) | Systems and methods for generating thin image slices from thick image slices | |
| CN111368849B (en) | Image processing method, image processing device, electronic equipment and storage medium | |
| Kudo et al. | Virtual thin slice: 3D conditional GAN-based super-resolution for CT slice interval | |
| US10867375B2 (en) | Forecasting images for image processing | |
| WO2014080961A1 (en) | Image processing device, image processing method and x-ray diagnosis device | |
| CN111369562B (en) | Image processing method, image processing device, electronic equipment and storage medium | |
| CN106530236B (en) | Medical image processing method and system | |
| CN114241077B (en) | CT image resolution optimization method and device | |
| JP6772123B2 (en) | Image processing equipment, image processing methods, image processing systems and programs | |
| JP2016189946A (en) | Medical image alignment device, method, and program | |
| CN108038840B (en) | Image processing method and device, image processing equipment and storage medium | |
| Li et al. | Learning non-local perfusion textures for high-quality computed tomography perfusion imaging | |
| Liu et al. | DL-MRI: A unified framework of deep learning-based MRI super resolution | |
| JP7106741B2 (en) | Learning method, learning device, generative model and program | |
| US11896362B2 (en) | Medical image processing apparatus, method of medical image processing, and nonvolatile computer readable storage medium storing therein medical image processing program | |
| Kerfoot et al. | Automated CNN-based reconstruction of short-axis cardiac MR sequence from real-time image data | |
| Shaikh et al. | Biomedical image processing and analysis using Markov random fields | |
| Na'am et al. | Enlarge medical image using line-column interpolation (LCI) method | |
| CN111862312A (en) | Brain blood vessel display device and method | |
| CN116740217B (en) | Arterial spin marking method, device and storage medium based on artificial intelligence technology |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| A621 | Written request for application examination |
Free format text: JAPANESE INTERMEDIATE CODE: A621 Effective date: 20161226 |
|
| A977 | Report on retrieval |
Free format text: JAPANESE INTERMEDIATE CODE: A971007 Effective date: 20171026 |
|
| A131 | Notification of reasons for refusal |
Free format text: JAPANESE INTERMEDIATE CODE: A131 Effective date: 20171114 |
|
| A521 | Request for written amendment filed |
Free format text: JAPANESE INTERMEDIATE CODE: A523 Effective date: 20171211 |
|
| A02 | Decision of refusal |
Free format text: JAPANESE INTERMEDIATE CODE: A02 Effective date: 20180109 |