GB2486496A - A method of identifying an object in an image by defining a series of bands - Google Patents

Abstract

The method extracts discriminating features of a candidate object in a CT, or other, image. Candidate regions are identified by a computer-aided diagnosis or computer-assisted detection (CAD) system, and are iteratively eroded from the surface to the core, generating a series of erosion bands. The erosion bands B1-B3 are then analyzed to extract features that differentiate between, for example, stool and polyps. The method may use, among others, a rule-based approach or an approach based on probability density functions of the erosion bands. The method can achieve significant reduction of false positives without any loss of sensitivity. The method can be applied, for example, to differentiate tagged or untagged stool from polyps in an image of a colon.

Description

Image Analysis

Field of the Invention

[00011 The present invention relates to image analysis. More particularly, the present invention relates to a method and apparatus for identifying a candidate object in a volumetric or 3D medical image.

Background of the Invention

[00021 Colorectal cancer is the second leading cause of cancer related death in westem countries and the third most common human malignancy in the United States. In recent years there has been much interest in Computed Tomography Colonography (CTC), which uses Computed Tomography (CT) images of a cleansed and insufflated colon. Computer Aided Detection (CAD) software is often used to assist a human expert (such as a radiographer, physician or oncologist) by automatically analysing the CT data and highlighting candidate objects as potential lesions. CAD may detect lesions that would have otherwise been missed by a human, and may give the human expert more confidence that lesions were not missed during the read. It has been demonstrated that the use of CAD leads to improved performance in detecting lesions in CTC data.

[00031 In CTC, stool may reduce the specificity by imitating polyps, and residual solid and liquid may reduce the sensitivity by obscuring the shape of polyps. In faecal-tagged CTC, orally administered tagging agents helps distinguish between stool and polyps, consequently improving specificity by opacThying residual fluid and stool. Figure 1 shows an example, in which Figure la is CT image of a polyp (P) and Figure lb is a CT image of tagged stool (TS). Without the contrast agent, it would be very difficult to differentiate between the two cases. Nonetheless, tagging also produces undesirable effects in CT intensity values (usually expressed in Hounsfield units, or HU) and when coating a polyp, can render the differentiation between polyps and stool very difficult. Tagged stool is one of the main sources of CAD false positives, followed by the ileoeecal valve, thick fold and extra-eolonie components.

[00041 Translucency rendering, which is a rendering of the candidate using ray tracing, has been used to detect the presence of tagged stool. Intensity values are accumulated until a maximum of opacity is reached along each ray's path. Accumulated values are codified using colours and opacities designed to differentiate polyps and different cases of false positives. The main shortcoming of this technique is that a good rendering requires robust computation of an axis perpendicular to the colon wall, which is a challenging task for the various shapes of polyps. Furthermore the characterization requires multiple complex thresholds for the colour and transparency used in the volume rendering transfer curve.

Statement of the Invention

[00051 According to the present invention, there is provided a method according to claim 1.

[0006] In embodiments of the invention, erosion-band histograms are used to detect tagged stool in medical images, and thereby avoid false positives caused by wrongly identifying tagged stool as a medical abnormality, such as a polyp. For a given candidate region, iterative erosions are performed from the candidate's border to its core, generating a series of bands. For each band, a histogram of intensity (HU values) is computed to characterize the intensities within the band, and features are extracted which are analyzed to differentiate between tagged stool and an abnormality such as a polyp. Erosion-band histograms provide a convenient and powerful mechanism for analyzing a 3D region using bands from its boundary to core. It is particularly useful in identifying tagged stool, since tagged stool often appears homogeneously bright and has high variance, while a polyp tends to have a darker core with lower variance.

[00071 In a first specific embodiment, a rule-based method classification system is used with rules formed by the statistics of intensities computed from each band in the series. In a second embodiment, kemel density estimation is used to produce probability density functions (PDFs), and K-nearest neighbour (KNIN) to classify a new candidate. Nearest neighbours are identified using information theoretic measures between a new candidate and a set of representative exemplars. In the third embodiment, each PDF extracted from the erosion-band histograms in sequence is concatenated to form a high dimensional feature vector, which is projected (linearly or nonlinearly) to a low dimensional subspace where a classifier is applied, for example to distinguish between a stool and a polyp.

[00081 Erosion-band histograms can be interpreted as probability density functions, which can then be compared using information theory (e.g. mutual information or Jensen-Shannon divergence (JSD)) to find nearest neighbours for kNN classification.

Brief Description of the Drawings

[00091 Specific embodiments of the invention will now be described, purely by way of example, with reference to the accompanying drawings in which: Figure 1 a is a CT image of a polyp; Figure lb is a CT image of tagged stool; Figure 2a is a 2D image of a polyp candidate object, with the outline of a mask defining the extent of the candidate object; Figure 2b shows the image of Figure 2a after applying a threshold; Figure 2c shows the image divided into erosion bands; Figure 3 is a flow diagram of a method according to a first embodiment; Figure 4 is a flow diagram of a method according to a second embodiment; Figure 5 is a graphical illustration of PDFs obtained for each erosion band of a candidate object, in the second embodiment; Figure 6 is a flow diagram of a method according to a third embodiment; Figure 7 is a graphical illustration of a single PDF obtained for each erosion band of a candidate object, in the third embodiment; Figures 8a and 8b are graphical illustrations of projections in 3D space for candidate objects using PCA and LE respectively.

Figure 9 is a schematic diagram showing a medical imaging device and a computer for processing image data from the medical imaging device; and Figure 10 is a more detailed diagram of the computer shown in Figure 9.

Detailed Description of Embodiments

Input Data [00101 Input data for specific embodiments of the invention comprise image data, preferably 3D or volumetric image data. However, other embodiments may operate on 2D data, using 2D versions of the 3D operations described below. The input image data may be digital or analog; in the latter case, the image is preferably digitised before being processed further.

[00111 Specific embodiments of the invention arc performed on a volumetric CT scan image which has been pre-processed by CAD software to identify one or more candidate objects within the image, for example by generating data defining the extent of the candidate objects. The candidate objects may have been selected as representing possible medical abnormalities, such as polyps. CAD methods of identifying such candidate objects are known, and may be based for example on intensity, differential geometric features such as shape index and curvedness, gradient concentration, texture, volume and other shape and/or binary features. An example of such methods is disclosed in A robust and fast system for CTC computer-aided detection of colorectal lesions', 0. Slabaugh, X. Yang, X. Ye, R. Boycs, and G. Beddoe, Algorithms, 3(1):21-43, 2010.

[00121 In one specific embodiment, the input data comprised 412 volumes from 8 different institutions, acquired by CT scanners with different scanning protocols, which had been pre-processed to identify 2048 candidate objects, of which 392 were polyps and 1656 were FPs.

The CAD system gave a performance of 90% sensitivity and 4.02 FPs per scan.

[00131 CAD-processed CT scans were used as the input to the embodiments of the invention, and the performance of the embodiments was measured on top of the CAD system's performance. For example, if the sensitivity of the embodiments was 100% with a 50% FP reduction, the overall CAD sensitivity is unchanged.

[00141 As an alternative, the candidate objects may be identified semi-automatically, for example by manual input by a user of a point within a candidate object, followed by automated identification of the extent of the candidate object, for example using region-growing. As another alternative, the candidate objects may be identified manually, by user input of a boundary containing the candidate object.

[0015] To study the performance of the embodiments on tagged stool, the CAD output was separated into tagged data and untagged data using an intensity threshold of 250 HU. If a candidate object has a single pixel with HU value greater or equal to 250, it is considered to be tagged data, otherwise as untagged data. Based on this threshold, the dataset consisted of 645 tagged candidates, including 139 candidates from polyps and 506 from FPs. For tagged data, HU values less than -64 are ignored so as to focus the analysis on soft tissue, bone and tagging material. The remaining 1403 candidates were categorized as untagged, including 253 polyps and 1150 FPs. For untagged data, HU values less than -439HU were ignored, as they are likely to correspond to air.

Erosion Bands [0016] In each of the specific embodiments, features of a 3D candidate object are captured using a banded approach. An objective is to characterize tagged stool and polyps, which may be achieved based on the intensity distribution of soft tissue, tagging materials, and air bubbles within the candidate object.

[00171 For tagged data, the method begins by applying a threshold of HU> -64 to exclude fat or air on the border of the candidate object, as shown in Figure 2. Figure 2a gives an example of a 2D view of a polyp candidate object, with the outline of a mask M0 defining the extent of the candidate object. In Figure 2b, the mask M after applying the threshold is illustrated in grey, and the border of the original mask discarded (HU < -64) is shown in slanting lines.

[00181 Next, 3D erosions are applied on the thresholded mask of Figure 2b to generate a series of erosion bands B1 from the candidate's surface to its core. Each erosion band provides a respective mask for extraction of image features to characterize the candidate object. An example is depicted in Figure 2c, where three erosion bands are rendered with different shading. Let M0 be the initial mask of a given candidate and F be a spherical structuring element. The series of erosion bands is defined as: I?= M1VM_ IvIi=IvIileE (1) where i ranges from [1,..., L] and L is the maximum number of erosions. V represents the XOR operator, and e is the morphological erosion operator. The i11, erosion band B1 is defined as the XOR operation between M and M11.

100191 3D erosion techniques are known per Se; they are described for example in "Tmage Processing and Mathematical Morphology: Fundamentals and Applications" by Frank Y. Shih, CRC Press, 2009. Other morphological operations may be used; for example, it is not essential that the morphological erosion operator be constant for each iteration; instead, the erosion bandwidth may vary by using varying erosion operators for each iteration.

[0020J The erosion bands may comprise layers or levels between distance transform isocontours. Alternative operations, such as distance transforms may be used to generate similar bands.

[00211 Figure 2c shows an example of a candidate of size 3, that is, L = 3. The HU values of the pixels masked within each erosion band B1 are used to perform the characterization of the candidate. This preserves the spatial distribution of the image intensities between the different bands, which is useful to distinguish a polyp surrounded by tagging and stool surrounded by tagging. This approach does not require any particular plane or projection as required by the prior art translucency approaches, and is robust to the variable 3D candidate object shape. Instead of solely or directly using intensity values within each band, this approach can be extended by using any other shape or texture information derived from the image. Non-limiting examples include the image gradient, shape index, curvedness, and gray-level co-occurrence features (GLCM). GLCM features are described in "Textural Features for Image Classification", Robert M Haralick, K Shanmugam, Its'hak Dinstein, IEEE Transactions on Systems, Man, and Cybernetics (1973) SMC-3 (6): 610-621.

[00221 Note that throughout the specification, the size of the candidate region is defined as the maximum number of erosions applied to reduce the region to the empty set. The candidate in Figure 2 has a size of L = 3.

[00231 In each of the following specific embodiments, a different approach is taken to extract information from the series of bands B1.

First Embodiment -Rule-based Approach [00241 The tagging agents used in CTC are designed to penetrate stool but not tissue.

Therefore, we expect the core of tagged stool candidate objects to be significantly brighter than polyp candidate objects. Additionally, the absorption of tagging agent may appear irregular for stool, leading to a higher variance of Hounsfield unit intensities.

[00251 In the first embodiment, a rule-based classification approach is used to identify tagged stool FPs. Two global features are first computed: volume and tagging percentage of the candidate object region. The percentage of tagging is a global measure of the presence of liquid in the candidate object, computed as a fraction of the volume with intensity above 130 HU. For each erosion band B1, a set of intensity-based statistics are computed including the standard deviation, minimum, maximum and median. With the two global measures and a set of statistics for each erosion band B1, some or all of the following rules are applied to differentiate tagged stool and polyp candidates: Ri: If the size of series for a candidate is 1 (L 1) and the median intensity value of B1 is above a threshold T1 (for example, 90 HU), then identify as FP. This rule aims to remove small stool.

R2: If the median intensity values of all B1 for a candidate object are above T2 (for example, 200 HU), then identify as FP. This rule aims to remove the candidate objects with regularly absorbed tagging.

R3: If the standard deviation in the last band BL for a candidate object is above T3 (for example, 130 HU), then identify as FP. This rule aims to remove the candidate objects with high variance, which is expected for tagged stool with the presence of air and tagging in its core.

R4: If the median intensity values of the last band BL for a candidate object is above T4 (for example, 400 HU), then identify as FP. This rule aims to remove candidate objects with strong tagging in their core, which are considered as tagged stool.

R5: If, in the last two erosion bands of a candidate, the median intensity value of B is less than the median of BLJ (i.e. median(BL) <median(Bi)), the candidate object is considered as a polyp candidate. This rule aims to capture the polyps surrounded by tagging and is used in conjunction with R6 and R7 to prevent identifying such polyps as FPs.

R6: If the median intensity values of each erosion band in the series monotonieally increase from B1 to BL, then identify as FP. This rule aims to remove tagged stool for which the intensity (HU value) increases towards its core from its border area due to absorbed tagging.

R7: If the percentage of tagging pixels in the candidate is above a threshold T6 (for example, 33%), where the tagging pixels are defined as voxels with intensity values above T7 (for example, 130 HU) as mentioned earlier, then identify as FP. This rule does not require erosion bands; it aims to remove those candidates with widespread tagging therein.

[00261 Some or all of the above-defined rules may comprise a rule-based classification system to remove FPs resulting from tagged stool, and may have a cascade implementation as shown by Figure 3. Candidate objects passing all the steps in the system arc considered as possible polyp candidates, otherwise they are identified as FPs. Note that this approach is restricted to tagged candidates and is not applied to untagged candidates.

Second Embodiment -Multiple PDF Approach PDF Generation [00271 The rule-based approach of the first embodiment described above is empirically designed and reliance on thresholds T for each rule requires a tuning stage. The second specific embodiment adopts a different approach that relies on the probability density function (PDF) associated with each band B1. The PDF gives a much more comprehensive characterization of the candidate object. Afterwards, a set of models are selected to perform a kNN-bascd classification. Figure 4 shows a flowchart of the process.

[00281 After the series of erosion bands of a candidate object is derived (S21), a histogram of intensity (HU) values is computed for each band. The underlying PDF of each histogram is then computed (S22), for example using a non-parametric kernel-smoothing density estimation method such as described in Applied Smoothing Techniques for Data Analysis: The Kernel Approach with S-Plus Illustrations', A. W. Bowman and A. Azzalini, Oxford Statistical Science Series, Oxford University Press, USA, November 1997. This method estimates a PDF of a set of independent and identically distributed random variables (xi, X2, xv). The density estimation is obtained by summing a set of kcmels distributed among these variables: n fhV =XK(Xh) (2) where K is the kernel used and h is the bandwidth which controls the amount of smoothing.

The normal kernel may be used: 1 12 K(x) (3) [00291 The bandwidth may be selected using the method proposed in Kernel density estimation via diffusion', Z. I. Botev, J. F. Grotowski, and D. P. Kroese, Annals of Statistics, 38(5), Nov 2010, which is a completely data-driven method that yields a good estimation even for multimodal densities. A PDF is estimated from the histogram of pixels values masked by each band Bi. A graphical example of the PDFs obtained is depicted in Figure 5, in which for each erosion band the abscissa axis represents the intensity values of the candidate, while the ordinate axis represents the estimated density.

Distance and Model Selection [00301 The distance between two candidate objects is defined using an information theoretic measure to compare the "distance" between corresponding bands of the objects: D(Cm, Ctm) = D( Cm(B3, Ctm(B3) (4) where D(C(B), C(B1)) represents the distance between candidates C' and C11 at band B1.

This requires both candidates to have the same number of bands (i.e. be the same size), where L represents the size of the candidate object, [B1, . . .BL]. The PDF from band B1 of a candidate object is compared with the PDF from the same number band B1 of the other candidate. For computational efficiency, the PDFs are sampled to have 400 bins. The distance between the two PDFs measures their difference, using mutual information (MI) or Jensen-Shannon divergence.

[00311 Mutual Information (MI) between two random variables X and Y (defined as each PDF in our case) is expressed as: p(xy) J(X;Y) = yEY XEX (5) where p(x, y) is the joint probability distribution and p(x), p(y) are the marginal probability distribution. This expression is also equivalent to: J(X;Y) = H(X,Y)-H(XIY)-H(YIX) (6) [00321 Jensen-Shannon divergence (JSD) is a symmetrized and smoothed version of Kullback-Leibler divergence measure based on information theory, calculated between two probability distributions P and Q, is expressed as: 1 1 JSD(P II Q) =DKL(P II M) +DKL(Q II M) (7) where P (i) DKL(P II M) = P(i) lo() (8) whereM =1(P+Q).

[00331 During training, a number of candidate objects representing polyp and tagged stool with different sizes are used as reference candidate objects, producing a set of models (S23).

The test candidate is compared to the models which have the same size. If the test candidate is closer to the models of polyps, it is classified as a polyp, othenvise as a non-polyp, or FP.

This is a 1-NN method. A kNN classifier may be used (S24) to take account of a larger neighbourhood: for example, when the majority of the k' closest neighbours to a candidate object are polyps, then this candidate would be classified as polyp. Otherwise, the candidate would be classified as non-polyp. More details of kNN classifiers are described in Pattern Classification (2nd Edition)', R. Duda, E. Hart and D. Stork, Wiley-Tnterscience, November, 2001.

Advantages and Disadvantages [00341 The PDF is a rich descriptor of the erosion band, and captures more information than a single statistic such as median or standard deviation. The limitation of this approach is that it requires distance between objects to be measured at each band. It makes it infeasible to have a single projection for a particular candidate object when working with a subspace method. Another limitation is the difficulty in finding good models for the kNN classifier.

Exhaustive search in a training dataset could be a computational burden and the result may not be optimal. In embodiments of the invention, random model selection or subspace model selection may be used; in the taller, a set of models are setected on the 3D subspace projected from a combined single PDF. The single PDF approach is described below in relation to the third embodiment.

Third Embodiment -Single PDF Approach [00351 An approach based on a single PDF to characterize the candidate may address the disadvantages of the second embodiment. Figure 6 shows a flow diagram of the method of the third embodiment, beginning with candidate object erosion (S3 1).

Single PDF [00361 Candidates of size one (L = 1) are characterized with a singte PDF that can be seen as a point in a D-dimensional space, where D' is the number of bins of the PDF. As D is typically large, it is difficult to visualize and classify points in a D dimensional space. Using a dimensionality reduction method such as PCA (Principle Component Analysis), a candidate in space IPI? can be projected to a subspace]R5' (e.g. S = 2; 3 i.e. two or three dimensions). However, in the present embodiment, a problem exists in that candidates of size (L> 1) consist of L PDFs. The PDFs must be combined to produce a singte vector for subspace projection. A solution such as summing or averaging the L PDFs does not preserve the spatial sequence used in computing the erosion bands and therefore would lose the spatial information.

[00371 Instead, in the present embodiment the PDFs of each band are concatenated together (S32) in a sequence from B1 to B. This buitds att L PDFs into a singte PDF to characterize a candidate by preserving all the information in the spatial order of erosions from surface to core. A graphical example is depicted in Fig. 7. Each single PDF can be represented by a point in a (L. D)-dimensional space for the candidates with same size L. PDF Projection [0038] The single PDF characterizes a candidate in a very high dimensional space.

Dimensionality reduction can be applied to the single PDF to simplify visualization and classification. In general, if we have two points x and x1 in a space, we need to find a proper distance measure which preserves the retationship between the points in its subspace.

[00391 Formatty, d(x,xj) = (x1 -xj)tA(xj -xj) = (A2x1 -Axj)t (A2x1 -A7x1) (9) where the projective matrix A is to be learned. A comprehensive survey of distance metric learning techniques can be found in Distance Metric Learning: A Comprehensive Survey', L. Yang and R. Jin. Technical report, Department of Computer Science and Engineering, Michigan State University, 2006.

[0040] In the training stage, for all the candidate objects with the same size k, the projective matrix A is learned, to project the data into an S-dimensional subspace. Repeating this process for different sizes, a set of projective matrices A, k E [1, L] is obtained for different candidate sizes. In the testing stage, we project the candidate data into an 5-dimensional subspace (S33) by applying the corresponding projective matrix 4.

[00411 Principal Component Analysis (PCA) is a linear projection method that preserves the global structure. Laplacian Eigenmaps (LE) is a non-linear technique -see Laplacian eigenmaps for dimensionality reduction and data representation', M. Belkin and P. Niyogi, Neural Computing, 15:1373-1396, June 2003. Figures 8a and 8b show graphical examples for candidates with size equal to 2 (L = 2). The projection obtained using PCA (Figure 8a) and LE (Figure 8b) successfully represents the data, showing a good separability between polyps (shown as white points) and FPs (shown as black points). However, other non-linear dimensionality reduction techniques can be applied, such as ISOMAP (see A Global Geometric Framework for Nonlinear Dimensionality Reduction', J. B. Tenenbaum, V. d.

Silva, and J. C. Langford, Science, 290(5500):2319-2323, 2000) or Locally Linear Embedding (LLE -see Nonlinear Dimensionality Reduction by Locally Linear Embedding', S. T. Roweis and L. K. Saul, Science, 290(5500):2323-2326, 2000).

[0042] After a representation of the candidate object is obtained in a 1It' subspace, a classifier such as Linear Discriminant Analysis or Naive Bayes classifier is applied (S34) to identify tagged stools. For further details of these techniques, see Statistical pattern recognition: a review', A. Jam, R. Duin, and J. Mao, Pattern Analysis and Machine Intelligence, IEEE Transactions on, 22(1):4 -37, Jan 2000.

[0043] The combined single PDF approach with a projection into its subspace and the subsequent classification of the data requires no empirical parameter settings to solve the problem in a principled manner.

Output Data [00441 As described above, the result of each of the specific embodiments is an identification of one or more candidate objects as being either a polyp or a false positive.

This identification may be a binary output, or a confidence level such as a percentage likelihood of the candidate object being a polyp.

[0045] The results may be output graphically, for example by highlighting candidate objects identified as polyps on a graphical display or print of the medical image and/or removing highlighting of candidate objects identified as false positives. Alternatively, the result may be a list of candidate objects identified as polyps and/or false positives, including references identifying each of the candidate objects. Alternatively, the output may be an alert that the medical image containing the one or more candidate objects has been identified as containing at least one polyp. Other alternatives may be apparent to the skilled person on

reading the description contained herein.

Experiments and Results First Embodiment [0046] Table 1 below shows the results obtained using the first embodiment for all datasets: Table 1 -Results for pt Embodiment Sensitivity Tagging FPS reduction Total FP reduction 98.75% 47.1% 15.06% Second Embodiment [00471 Table 2 below shows the results obtained using the second embodiment, as an average for 100 repetitions for each percentage of the data used as randomly selected models: TabJe 2 -Results for 2'" Embodiment Cross-validation for KNN classification Percent Sens. (%) Std FP Reduction (%) Std 75.1 3.3 93.6 1.2 86.9 2.1 96.2 0.7 [00481 JSD showed the best performance as distance measure in the experiments, and MI ranks as second; both are better than Euclidean distance. Table 2 gives the cross-validation performance of JSD with 1KNN method (K==12). For this experiment, a certain percentage of the tagging data was selected to be used as models. The rest of the tagging data is classified using the JSD value between the rest of the dataset and the models following a kNN strategy. For a new candidate, the JSD is computed between the PDFs associated to its BL and those of each model. The majority of K lowest distance models representing polyps or FPs will determine which class the new candidate object belongs to. This experiment has been performed 100 times when 50% and 70% data are used as models respectively. The low standard deviation points out the stability of this approach.

[0049] However, random selection of models generally cannot produce the best models for the classifier. To make the model selection more targeted to the discriminative models, the single PDF method may be used to project into its 3-dimensional subspace for data with each size, as explained above. In this 3-dimensional subspace, a Naive Bayes classifier is rnn, and the first N closest candidate objects to the decision boundary are selected (see Figure 8). This approach is not just limited to a specific decision boundary and also allows selection of some outliers. Using this approach, a total of 23 FP models and 19 polyp models are selected in the tagged dataset, representing 6.5% of the overall data. With these models a kNN classification is run, obtaining a 52% FP reduction without any loss of sensitivity. Note however, that model selection in this way makes the multiple PDF approach dependent on the single PDF approach.

Third Embodiment [0050] To demonstrate the performance of the single PDF approach of the third embodiment, the following four experiments were performed. The first three experiments are performed on tagged data only.

[0051] In the first experiment, a single PDF approach was applied as described above. Data of all sizes was combined by zero-padding the single PDF so that a single projective matrix for the whole data can be learned in the training, avoiding the issue of different candidate sizes in the classification. The classifier was trained with 45% of the data and independently tested on the rest. Table 3 below shows the results of Naive Bayes using a 4D subspaee projected by PCA and a 1OD subspace projected by LE respectively. It can achieve 32.2% FP reduction in the tagged data without the loss of CAD sensitivity.

Table 3 -FP Reduction at different sensitivities obtained using tagged data and a single projective matrix A for all candidate object sizes FP Reduction Sensitivity (%) 99.3 98.56 96.4 PCA4D 32.2% 32.6% 36.56% 43.48% LE 1OD 25.3% 40.9% 41.7% 43.48% [00521 In the second experiment, the candidate size was taken into account. A set of projective matrices Ak was learned independently from the data at each size L, and a Naive Bayes classifier was applied independently for each subspaee. This achieved 54% of FPs reduction without any loss of sensitivity in the tagged data. This result was obtained using LE and projecting the data into a 3D subspace. With PCA, the reduction of FPs was 49.4%.

[00531 In the third experiment, the whole candidate region was taken and its PDF extracted directly without any erosion. The PDF was projected into its subspace and Naive Bayes was applied to classify the data on the subspace. FPs were reduced by 19.34% without the loss of CAD sensitivity in the tagged data. This performance is inferior to the approach with a series of PDFs estimated from the bands, which shows the importance of keeping image intensity information in order of its spatial distribution.

100541 In the fourth experiment, the approach was applied to untagged data. This data includes fat, as well as soft tissue, with HU values between -439 and 250. When the classification was performed as in the second approach, 17.8% FP reduction was achieved without impact on CAD sensitivity.

100551 Combining the second and fourth experiment, the results are summarized in Table 4 below.

Table 4 -Global results for tagged and untagged data Sensitivity(%) 100 99.5 98.7 98 FPReduction(%) 28.9 31.82 33.33 35.3 [0056] Table 4 shows that the single PDF method is able to reduce 28.9% of total FPs without any loss of sensitivity. This means that the average, overall FPs per scan is reduced to 2.86 which is a significant improvement of the CAD system.

[0057] The distinction between tagged data and the untagged data is important to achieve the best results. Mixing them together complicates the data distribution and yields worse results.

Conclusions

[0058] The above embodiments demonstrate that a series of spatially-related PDFs or statistics extracted from a set of erosion bands can distinguish tagged EPs and polyps in a CAD system. Three different specific embodiments employ different methods to extract new features from the set of erosion bands to differentiate tagged stool from polyps.

Experimental results show that each method can reduce tagged FPs significantly and the best method based on the single PDF with Naive Bayes classifier can reduce total FPs by 28.9% without any loss of sensitivity. These results show improvement in the performance of CAD and therefore confidence of radiologists in a CAD system.

Apparatus 100591 An example of the apparatus used to implement the invention will now be described with reference to Figures 9 and 10. As shown in Figure 9, a scan image may be created by a computer 504, which receives scan data from a scanner 502 and constructs the scan image.

The scan image is saved as an electronic file or a series of files, which are stored on a storage medium 506, such as a fixed or removable disc. The scan image may include metadata associated with the scan image. The scan image may be analysed by the computer 504, or the scan image may be transferred to another computer 508 which runs software for processing the scan image in order to perform the method described above. The software may be stored on a carrier, such as a removable disc or a solid-state memory, or downloaded over a network such as a local area network (LAN), wide-area network (WAN), an internet or the Internet.

[00601 The computers described herein may be computer systems 600 as shown in Figure 10. Embodiments of the present invention may be implemented as programmable code for execution by the computer system 600. Various embodiments of the invention are described in terms of this example computer system 600. After reading this description, it will become apparent to a person skilled in the art how to implement the invention using other computer systems and/or computer architectures.

[00611 Computer system 600 includes one or more processors, such as processor 604.

Processor 604 may be any type of processor, including but not limited to a special purpose or a general-purpose digital signal processor. Processor 604 is connected to a communication infrastructure 606 (for example, a bus or network). Computer system 600 also includes a main memory 608, preferably random access memory (RAM), and may also include a secondary memory 610. Secondary memory 610 may include, for example, a hard disk drive 612 and/or a removable storage drive 614, representing a floppy disk drive, a magnetic tape drive, an optical disk drive, etc. Removable storage drive 614 reads from and/or writes to a removable storage unit 618 in a well-known manner. Removable storage unit 618 represents a floppy disk, magnetic tape, optical disk, etc., which is read by and written to by removable storage drive 614. As will be appreciated, removable storage unit 618 includes a computer usable storage medium having stored therein computer software and/or data.

[0062] In alternative implementations, secondary memory 610 may include other similar means for allowing computer programs or other instructions to be loaded into computer system 600. Such means may include, for example, a removable storage unit 622 and an interface 620. Examples of such means may include a program cartridge and cartridge interface (such as that previously found in video game devices), a removable memory chip (such as an EPROM, or PROM, or flash memory) and associated socket, and other removable storage units 622 and interfaces 620 which allow software and data to be transferred from removable storage unit 622 to computer system 600. Alternatively, the program may be executed and/or the data accessed from the removable storage unit 622, using the processor 604 of the computer system 600.

[0063] Computer system 600 may also include a communication interface 624.

Communication interface 624 allows software and data to be transferred between computer system 600 and external devices. Examples of communication interface 624 may include a modem, a network interface (such as an Ethernet card), a communication port, a Personal Computer Memory Card International Association (PCMCIA) slot and card, etc. Software and data transferred via communication interface 624 are in the form of signals 628, which may be electronic, electromagnetic, optical, or other signals capable of being received by communication interface 624. These signals 628 are provided to communication interface 624 via a communication path 626. Communication path 626 carries signals 628 and may be implemented using wire or cable, fibre optics, a phone line, a wireless link, a cellular phone link, a radio frequency link, or any other suitable communication channel. For instance, communication path 626 may be implemented using a combination of channels.

[0064] The terms "computer program medium" and "computer usable medium" are used generally to refer to media such as removable storage drive 614, a hard disk installed in hard disk drive 612, and signals 628. These computer program products are means for providing software to computer system 600. However, these terms may also include signals (such as electrical, optical or electromagnetic signals) that embody the computer program disclosed herein.

[0065] Computer programs (also called computer control logic) are stored in main memory 608 and/or secondary memory 610. Computer programs may also be received via communication interface 624. Such computer programs, when executed, enable computer system 600 to implement the present invention as discussed herein. Accordingly, such computer programs represent controllers of computer system 600. Where the invention is implemented using software, the software may be stored in a computer program product and loaded into computer system 600 using removable storage drive 614, hard disk drive 612, or communication interface 624, to provide some examples.

[0066] In alternative embodiments, the invention can be implemented as control logic in hardware, firmware, or software or any combination thereof The apparatus may be implemented by dedicated hardware, such as one or more application-specific integrated circuits (ASICs) or appropriately connected discrete logic gates. A suitable hardware description language can be used to implement the method described herein with dedicated hardware.

Alternative Embodiments [0067] The embodiments described above are illustrative of rather than limiting to the present invention. Alternative embodiments apparent on reading the above description may nevertheless fall within the scope of the invention.

Claims (33)

1. Claims 1. A method of identifying a candidate object in an image thereof, the method comprising: a. performing morphological operations on the image to define a series of bands of the object; b. deriving one or more features of each of the series of bands; and c. classifying the features so as to identify the candidate object.
2. 2. The method of claim 1, wherein the morphological operations comprise a plurality of iterative morphological operations.
3. 3. The method of claim 1 or claim 2, wherein the operations comprise successive erosions of the candidate object.
4. 4. The method of any preceding claim, including, prior to step a, thresholding the image to exclude one or more portions of the image bordering the candidate object and not forming a part thereof
5. 5. The method of any preceding claim, wherein the features are derived from image intensities of each of the series of bands.
6. 6. The method of any preceding claim, wherein the features are derived from shape attributes of the series of bands.
7. 7. The method of claim 6, wherein the shape attributes comprise one or more of shape index and curvedness.
8. 8. The method of any preceding claim, wherein the features are derived from textural attributes of the series of bands.
9. 9. The method of any preceding claim, wherein the features are derived from one or more of image gradient and gray-level co-occurrence.
10. 10. The method of any preceding claim, wherein the features comprise statistical measures of each of the series of bands.
11. 11. The method of any preceding claim, wherein step c comprises applying one or more rules based on the features.
12. 12. The method of any one of claims 1 to 9, wherein the features comprise probability density functions of each of the bands.
13. 13. The method of claim 12, wherein the probability density functions are derived from histograms of image intensities within each band.
14. 14. The method of claim 12 or claim 13, wherein step c comprises comparing the probability density functions of the candidate object with those of one or more reference objects.
15. 15. The method of claim 14, comprising deriving a distance between corresponding bands of the candidate object and of the one or more reference objects.
16. 16. The method of claim 15, wherein the distance comprises Jensen-Shannon divergence, Mutual Information, Mahalanobis distance or Euclidean distance.
17. 17. The method of claim 14, wherein the probability densities of each of the bands are combined to form a single probability density function for the candidate object.
18. 18. The method of claim 17, wherein the dimensions of each of the single probability density functions are reduced before classification.
19. 19. The method of claim 18, wherein the dimensions are reduced using principle component analysis (PCA) or Laplacian Eigenmaps (LE).
20. 20. The method of any one of claims 17 to 19, wherein classification comprises one or more of Naïve Bayes, Support Vector Machine (SVM), Linear Discriminant Analysis and Quadratic Discriminant Analysis.
21. 21. The method of any preceding claim, wherein the image comprises a volumetric image and the morphological operations comprise volumetric operations.
22. 22. The method of any preceding claim, wherein the image comprises a medical image.
23. 23. The method of claim 22, wherein the medical image is a CT image.
24. 24. The method of claim 22 or 23, wherein the medical image is of a part of a colon.
25. 25. The method of any one of claims 22 to 24, wherein the candidate object is a potential polyp.
26. 26. The method of any one of claims 22 to 25, wherein the image is of a faecal-tagged part of a colon.
27. 27. The method of any preceding claim, comprising outputting an identification of the candidate object.
28. 28. A method substantially as herein disclosed with reference to and/or as shown in the accompanying drawings.
29. 29. Apparatus arranged to perform the method of any preceding claim.
30. 30. A computer-readable medium comprising instructions which, when executed by a suitably arranged computer or the like, cause the same to perform the method of any one of claims ito 28.
31. 31. The medium of claim 30, comprising a non-transient medium.
32. 32. The medium of claim 30, comprising a signal.
33. 33. Candidate object identification data produced by the method of any one of claims i to 28.