CN111402269A - Vertebral canal segmentation method based on improved FC-DenseNuts - Google Patents
Vertebral canal segmentation method based on improved FC-DenseNuts Download PDFInfo
- Publication number
- CN111402269A CN111402269A CN202010186436.XA CN202010186436A CN111402269A CN 111402269 A CN111402269 A CN 111402269A CN 202010186436 A CN202010186436 A CN 202010186436A CN 111402269 A CN111402269 A CN 111402269A
- Authority
- CN
- China
- Prior art keywords
- improved
- densenets
- network
- image
- data set
- 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
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10072—Tomographic images
- G06T2207/10081—Computed x-ray tomography [CT]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30008—Bone
- G06T2207/30012—Spine; Backbone
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
Abstract
A vertebral canal segmentation method based on improved FC-Densenets belongs to the deep learning medical image segmentation direction and comprises the following steps of performing data enhancement on a CT image data set of a vertebral part to obtain an enhanced CT image data set, dividing the enhanced CT image data set to obtain a training set and a testing set, performing feature extraction on the enhanced training set by adopting a target detection YO L O V3 algorithm to determine a target area, obtaining an improved FC-Densenets network I by adding long-short jump connection to the FC-Densenets network, performing dimension reduction convolution kernel operation on the improved FC-Densenets network I to obtain an improved FC-Densenets network structure diagram, training the improved FC-Densenets network, obtaining a vertebral part segmentation effect diagram based on the testing set, and obtaining the deep learning network by adopting the improved FC-Densenets Fully structure diagram to improve the effect.
Description
Technical Field
The invention relates to a deep learning medical image segmentation direction, in particular to a vertebral canal segmentation method based on improved FC-DenseNuts.
Background
In the present society, doctors are always lacking in the society, huge medical needs exist in some remote areas, the workload of doctors is very large, the diagnosis of the spondylopathy always depends on the need of manual detection, and therefore, the processing of medical images is generated, and the objective implementation and subjective dependence are hoped to be solved.
Conventional image segmentation methods include structure-based methods, grayscale-based methods, and texture-based methods. Also, a structure-based method is proposed to represent the shape characteristics of the region of interest by constructing a probability model, and the segmented region is continuously updated under the constraint of the shape model, so as to obtain an accurate segmentation result, wherein the Koh J first proposes a brand-new algorithm, the spinal canal can be automatically extracted unsupervised based on a significant attention model and a standard active contour method, and then the spinal cord is detected by using a dynamic programming algorithm.
S utilizes level set, Gaussian filtering and two morphological operations to segment the vertebral body, and finally, soft tissue structures such as spinal cord and capsule are arranged in the lumbar vertebra area, so that the segmentation complexity is increased. On the other hand, Wang Q uses a random walk method to enhance the segmentation result domain to achieve an average Dice accuracy of 95.61 ± 2.25%, however, the vertebral body is mainly composed of bone, and soft tissue structures such as spinal cord and capsule are present inside the vertebral canal, so the segmentation complexity is also increased.
Another relatively novel Convolution mode is the hole Convolution (scaled Convolution) proposed by Yu and Fisher, the main mechanism of the hole Convolution is to convolute with a larger Convolution range along with a certain sampling interval, so as to obtain a feature map with a wider visual field, and a plurality of sampling rates are combined to obtain multi-scale Context (Multiple Context) features.
However, these methods all have respective defects, the conventional methods generally have poor anti-interference capability and no generalization in application, and for example, for a method based on a deep learning framework, if a network model is too complex, although the dige is improved, the parameters are too large, the spatial dimension is too high, and the requirements of real-time performance can not be met, the real-time performance is that the whole process from the acquisition of medical images to the completion of processing is optimized in a shorter time, the rapid response process can be maintained only by reducing the parameter amount of the algorithm, simplifying the structure and reducing the running memory, and meanwhile, the method has a better application value in order to adapt to the real-time access of an intelligent mobile terminal.
Disclosure of Invention
According to the problems existing in the prior art, the invention discloses a vertebral canal segmentation method based on improved FC-DenseNuts, 1, the vertebral canal segmentation method based on the improved FC-DenseNuts comprises the following steps:
s1: performing data enhancement on the CT image data set of the vertebra part to obtain an enhanced CT image data set, and dividing the enhanced CT image data set to obtain a training set and a test set;
s2, adopting a target detection YO L O V3 algorithm to extract the features of the enhanced training set and determine a target area;
s3: obtaining an improved FC-Densenets network I by adding long and short hop connection to the FC-Densenets network, and performing dimensionality reduction convolution kernel operation on the improved FC-Densenets network I to obtain an improved FC-Densenets network structure diagram;
and S4, training the improved FC-transmissions network structure diagram, and obtaining a segmentation effect diagram of the vertebra part based on the test set.
Further: enhancing the CT image dataset includes: rotation, horizontal flipping, cropping, resizing, and adding image noise.
Further, feature extraction is carried out on the enhanced CT image data set by adopting a target detection YO L O V3 algorithm, and a specific process for determining a target region is as follows, an anchor frame of the CT image is determined through size clustering, width, height and midpoint coordinates are predicted for each frame network, and a vertebral canal image is extracted to obtain an integral feature region.
Due to the adoption of the technical scheme, the vertebral canal segmentation method based on the improved FC-DenseNuts can effectively solve the problems of errors and uncertain factors of the subjective judgment of a doctor on an image in the actual medical diagnosis process; under the condition that the region to be segmented occupies a small proportion of the whole CT image, the target detection network is adopted to extract the whole features of the image; the deep learning network adopts an improved FC-DenseNutsFully (connected DenseNuts) structure network to improve the segmentation effect, and has less parameters, lower running memory and a simplified network structure after improvement, so that the corresponding real-time performance is improved, and the anti-interference performance is obvious.
Drawings
In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings needed to be used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments described in the present application, and other drawings can be obtained by those skilled in the art without creative efforts.
FIG. 1 is a schematic process diagram of the process of the present invention;
FIG. 2 is a detailed network structure diagram of the improved FC-DenseNuts;
FIG. 3(a) is a less characteristic CT image;
FIG. 3(b) is a characteristic CT map;
FIG. 4(a) is a standard Mask template diagram;
FIG. 4(b) is a data set label diagram;
fig. 4(c) is a diagram showing the result of FC-DenseNets network segmentation.
Detailed Description
In order to make the technical solutions and advantages of the present invention clearer, the following describes the technical solutions in the embodiments of the present invention clearly and completely with reference to the drawings in the embodiments of the present invention:
FIG. 1 is a schematic process diagram of the process of the present invention; a method of spinal segmentation based on improved FC-DenseNets comprising the steps of:
s1, converting the acquired medical vertebra CT image format into a required common image format, increasing data volume in a classification mode to form a data set to obtain an enhanced CT image data set, and dividing the enhanced CT image data set to obtain a training set and a testing set; marking out a vertebral canal region to be segmented in the data set, and manufacturing the data set containing a mask label of the vertebral canal region according to a standard data set format;
s2, adopting a target detection YO L O V3 algorithm to extract the characteristics of the data set obtained in the S1, mainly determining an anchor box of the CT image through size clustering, and predicting 4 coordinates t of each bounding box networkx，ty，tw，thIn the algorithm bx、by、bwAnd bhI.e. the predicted coordinates and size of the center of the bounding box, cxAnd cyAnd extracting an integral characteristic region by adopting fusion of a plurality of scales for further segmentation processing. Wherein:
bx＝σ(tx)+cx(1)
by＝σ(ty)+cy(2)
s3: obtaining an improved FC-Densenets network I by adding a long-short jump connection mode to the FC-Densenets network, carrying out deep extraction on the characteristics of a target area, and carrying out dimensionality reduction convolution kernel operation on the improved FC-Densenets network I to obtain an improved FC-Densenets network structure diagram;
the specific network structure of the invention is constructed by adopting the characteristics of U-type coding and decoding and dense connection of the FC-DenseNuts algorithm, fig. 2 is a specific network structure diagram of the improved FC-DenseNuts, and the specific method is as follows for the improved FC-DenseNuts network training: the method comprises the steps of generating a picture path list, a training list and a verification list from a data set through a python script file, modifying json files corresponding to the data set respectively, operating a command line to load an initial weight file to train a network, and performing a training test process, wherein the encoding characteristics are superposed on decoding characteristic layers in the improved process for an original algorithm when resolution is restored, and the Concat process directly influences the extraction of shallow characteristics, so that the improved measure is to increase the skip-connection link of the improved process to fully extract the encoding process. The spatial dimension is higher and higher in the process of featuremap superposition, and the improved measure is that the 1 × 1 convolution kernel structure is subjected to dimension reduction processing. The sense field of each element in the feature map can be expanded by combining a target detection network for feature loss in the segmentation process, wherein the DenseBlock link is included in which the input of each layer is connected together by using the output of the previous layer and the input of the previous layer as the input of the layer.
Merging channels by using feature maps of the outputs of the front 0 to L layers;
xl＝1-Hl([x0，x1，…，xl-1，]) (4)
meanwhile, for the data set in S1, a CT image with unobvious characteristics can be obtained in FIG. 3(a), a CT image with obvious characteristics in FIG. 3(b), if the traditional image processing technology is adopted, the anti-interference performance is particularly poor, the adopted Bemsens self-adaptive local segmentation algorithm is used for test comparison, the Bernsens algorithm is a typical self-adaptive local segmentation algorithm, the average value of the maximum value and the minimum value of the gray level of each pixel in a window is used as the threshold value of the central pixel of the window, therefore, the method has no preset threshold value and has wider adaptability compared with the whole threshold value method, is not influenced by the conditions of non-uniform illumination and the like, and is used for anti-interference test, the gray value f (i, j) of the image at the pixel point (i, j) is set in the algorithm, the window is considered, the window (2 omega +1) × (2 omega +1) with the pixel point (i, j) as the center, wherein (2 omega +1) represents the side length of the window, the Bernsen algorithm can be described as calculating the threshold value (i, j) of each pixel point (i, j) in the,
and carrying out point-by-point binarization on each pixel point (i, j) in the image by using the value of b (i, j).
S4: training the improved FC-densitenes network structure diagram, and obtaining a segmentation effect diagram of the vertebra part based on the test set. Comparing the over-segmentation and under-segmentation conditions under the original data sample label, and FIG. 4(a) is a standard Mask template diagram; fig. 4(b) is a data set label diagram, and fig. 4(c) is a diagram of FC-DenseNets network segmentation result.
Further, the data enhancement is performed by converting the DICOMD picture, and the method for enhancing the CT image data set includes: rotating, horizontally turning, shearing, changing size, increasing image noise, expanding data volume, labeling, and creating data set according to standard data format.
Further, feature extraction is carried out on the enhanced CT image data set by adopting a target detection YO L O V3 algorithm, and a target region is determined by the specific process that an anchor frame of the CT image is determined through size clustering, width, height and midpoint coordinates are predicted for each frame network, and a vertebral canal image is extracted to obtain an integral feature region.
Furthermore, a mask result template obtained by testing is used for extracting a target area on the original image of the medical image, and difference labeling is carried out on the corresponding insufficiently-segmented places, so that the characteristics of rapidness, accuracy and strong anti-interference performance compared with other algorithms can be obviously seen from the obtained final result.
The above description is only for the preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art should be considered to be within the technical scope of the present invention, and the technical solutions and the inventive concepts thereof according to the present invention should be equivalent or changed within the scope of the present invention.
Claims (3)
1. A vertebral canal segmentation method based on improved FC-DenseNuts is characterized in that: the method comprises the following steps:
s1: performing data enhancement on the CT image data set of the vertebra part to obtain an enhanced CT image data set, and dividing the enhanced CT image data set to obtain a training set and a test set;
s2, adopting a target detection YO L O V3 algorithm to extract the features of the enhanced training set and determine a target area;
s3: obtaining an improved FC-Densenets network I by adding long and short hop connection to the FC-Densenets network, and performing dimensionality reduction convolution kernel operation on the improved FC-Densenets network I to obtain an improved FC-Densenets network structure diagram;
and S4, training the improved FC-transmissions network structure diagram, and obtaining a segmentation effect diagram of the vertebra part based on the test set.
2. The improved FC-DenseNets based spinal segmentation method according to claim 1, wherein: enhancing the CT image dataset includes: rotation, horizontal flipping, cropping, resizing, and adding image noise.
3. The improved FC-DenseNuts-based spine segmentation method according to claim 1, wherein a target detection YO L O V3 algorithm is adopted to perform feature extraction on the enhanced CT image data set, and the specific process of determining the target region comprises the steps of determining an anchor frame of the CT image through size clustering, predicting width, height and midpoint coordinates of each frame network, and extracting the spine image to obtain an overall feature region.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010186436.XA CN111402269A (en) | 2020-03-17 | 2020-03-17 | Vertebral canal segmentation method based on improved FC-DenseNuts |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010186436.XA CN111402269A (en) | 2020-03-17 | 2020-03-17 | Vertebral canal segmentation method based on improved FC-DenseNuts |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN111402269A true CN111402269A (en) | 2020-07-10 |
Family
ID=71434331
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010186436.XA Pending CN111402269A (en) | 2020-03-17 | 2020-03-17 | Vertebral canal segmentation method based on improved FC-DenseNuts |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111402269A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2022037548A1 (en) * | 2020-08-17 | 2022-02-24 | 浙江大学 | Mri spinal image keypoint detection method based on deep learning |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109035184A (en) * | 2018-06-08 | 2018-12-18 | 西北工业大学 | A kind of intensive connection method based on the deformable convolution of unit |
| CN109559295A (en) * | 2018-06-04 | 2019-04-02 | 新影智能科技(昆山)有限公司 | Image analysis system, method, computer readable storage medium and electric terminal |
| CN110415291A (en) * | 2019-08-07 | 2019-11-05 | 清华大学 | Image processing method and relevant device |
-
2020
- 2020-03-17 CN CN202010186436.XA patent/CN111402269A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109559295A (en) * | 2018-06-04 | 2019-04-02 | 新影智能科技(昆山)有限公司 | Image analysis system, method, computer readable storage medium and electric terminal |
| CN109035184A (en) * | 2018-06-08 | 2018-12-18 | 西北工业大学 | A kind of intensive connection method based on the deformable convolution of unit |
| CN110415291A (en) * | 2019-08-07 | 2019-11-05 | 清华大学 | Image processing method and relevant device |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2022037548A1 (en) * | 2020-08-17 | 2022-02-24 | 浙江大学 | Mri spinal image keypoint detection method based on deep learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110490212B (en) | Molybdenum target image processing equipment, method and device | |
| CN106940816B (en) | CT image pulmonary nodule detection system based on 3D full convolution neural network | |
| CN109840483B (en) | Landslide crack detection and identification method and device | |
| CN109389585B (en) | Brain tissue extraction method based on full convolution neural network | |
| Zhang et al. | Intelligent scanning: Automated standard plane selection and biometric measurement of early gestational sac in routine ultrasound examination | |
| CN111414923B (en) | Indoor scene three-dimensional reconstruction method and system based on single RGB image | |
| CN110570350A (en) | two-dimensional follicle detection method and device, ultrasonic equipment and readable storage medium | |
| CN112446892A (en) | Cell nucleus segmentation method based on attention learning | |
| CN112365973B (en) | Pulmonary nodule auxiliary diagnosis system based on countermeasure network and fast R-CNN | |
| CN112132854B (en) | Image segmentation method and device and electronic equipment | |
| CN112991365B (en) | Coronary artery segmentation method, system and storage medium | |
| CN111369574A (en) | Thoracic cavity organ segmentation method and device | |
| CN111402254A (en) | CT image pulmonary nodule high-performance automatic detection method and device | |
| CN112950561A (en) | Optical fiber end face defect detection method, device and storage medium | |
| CN112308822A (en) | Intervertebral disc CT image detection method based on deep convolutional neural network | |
| CN113077419A (en) | Information processing method and device for hip joint CT image recognition | |
| CN115984555A (en) | Coronary artery stenosis identification method based on depth self-encoder composition | |
| CN103169506A (en) | Ultrasonic diagnosis device and method capable of recognizing liver cancer automatically | |
| CN115100494A (en) | Identification method, device and equipment of focus image and readable storage medium | |
| CN114299080A (en) | Throat organ segmentation method based on cavity residual error characteristic pyramid | |
| CN111402269A (en) | Vertebral canal segmentation method based on improved FC-DenseNuts | |
| CN111462084B (en) | Image vectorization printing bleeding point prediction system and method based on random forest | |
| CN117523203A (en) | Image segmentation and recognition method for honeycomb lung disease kitchen based on transducer semi-supervised algorithm | |
| CN112862786B (en) | CTA image data processing method, device and storage medium | |
| CN112862785B (en) | CTA image data identification method, device and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |