Breast Cancer is the most prevalent cancer among women across the globe. Automatic detection of breast cancer using Computer Aided Diagnosis (CAD) system suffers from false positives (FPs). Thus, reduction of FP is one of the challenging tasks to improve the performance of the diagnosis systems. In the present work, new FP reduction technique has been proposed for breast cancer diagnosis. It is based on appropriate integration of preprocessing, Self-organizing map (SOM) clustering, region of interest (ROI) extraction, and FP reduction. In preprocessing, contrast enhancement of mammograms has been achieved using Local Entropy Maximization algorithm. The unsupervised SOM clusters an image into number of segments to identify the cancerous region and extracts tumor regions (i.e., ROIs). However, it also detects some FPs which affects the efficiency of the algorithm. Therefore, to reduce the FPs, the output of the SOM is given to the FP reduction step which is aimed to classify the extracted ROIs into normal and abnormal class. FP reduction consists of feature mining from the ROIs using proposed local sparse curvelet coefficients followed by classification using artificial neural network (ANN). The performance of proposed algorithm has been validated using the local datasets as TMCH (Tata Memorial Cancer Hospital) and publicly available MIAS (Suckling et al., 1994) and DDSM (Heath et al., 2000) database. The proposed technique results in reduction of FPs from 0.85 to 0.02 FP/image for MIAS, 4.81 to 0.16 FP/image for DDSM, and 2.32 to 0.05 FP/image for TMCH reflecting huge improvement in classification of mammograms.
Breast cancer is the most common cancer disease among women across worldwide. It is the leading cause of deaths for women suffering from cancer disease in India. It is estimated that breast cancer cases in India would reach to as high as 1,797,900 by 2020 [
The design and development of CAD system is an important progressive area of research for contrast enhancement for better visualization and clarification [
Several researchers have implemented clustering method like K-means and Fuzzy C-means (FCM) for breast abnormality segmentation [
In this work, novel method of extracting sparse curvelet subband coefficients by incorporating the knowledge of irregular shape of masses as they appear in sparse matrix and calculating LBP features has been presented. Therefore, this paper presents scheme as follows: Preprocessing of mammogram image for contrast enhancement using local entropy maximization-based image fusion algorithm and removal of background noise Cluster-based segmentation of mammograms using SOM and extract tumor regions, i.e., ROI) FP reduction: extraction of sparse curvelet subband coefficients and computation of LBP descriptor to classify true positives and false positives to improve performance of CAD system using MIAS [
The organization of paper is as follows: Sections
The block schematic of proposed integrated method for automatic detection of breast cancer using sparse curvelet coefficient-based LBP descriptor has been shown in Figure
Schematic architecture for automatic breast cancer detection.
The mammogram images are low-dose x-ray images so they have poor contrast and suffer from noises. The preprocessed mammogram image as shown in Figures
Steps for mammogram processing (a) enhanced mammogram, (b) binary mask, (c) pectoral removal, (d) pectoral removed mammogram, (e) clustered image, (f) cluster of interest, and (g) ROI extraction.
The contrast enhancement of the mammogram is performed using local entropy maximization [
Load input image (img1) Apply CLAHE algorithm and obtain enhanced image (img2) Decompose img1 and img2 up to 3 level of decomposition using Discrete Wavelet transform (DWT) Use maximum local entropy rule for fusion of img1 and img2 for high frequency subbands Take inverse DWT to obtain the fused image
Preprocessing. (a) Original image from MIAS database. (b) Contrast-enhanced mammogram using local entropy maximization. (c) Process of pectoral muscle removal. (d) Pectoral muscle removed mammogram.
Pectoral muscle suppression has been performed by defining rectangle as suggested in [
SOM is a special type of neural network designed to map the input image of size
At the start, weight vector
Weight vector for winning output neuron and its neighboring neurons are updated as
Figures
FP reduction by thresholding (a) clustered image, (b) clusters boundaries marked on original image, and (c) clusters after thresholding.
We can see that there are many FPs along with TP (marked by pink color) which are reduced using pixel level threshold (PLT based on pixel count in TP) as explained above. Figure
After SOM clustering (initial segmentation), the next step is to classify the detected regions into TP and FP by using proposed local sparse curvelet features (LSCF) followed by ANN classifier. To do so, initially, we have extracted ROIs from detected regions by SOM clustering and manually categorized into TP and FP. We collected these ROIs from three different datasets according to their maximum height and maximum width using connected components e.g., region marked in Figure
Variable sizes ROIs from MIAS, DDSM, and TMCH datasets.
After ROI extraction, FP reduction algorithm performs computation of proposed local sparse curvelet features (LSCF) followed by ANN classifier.
LBP [
Computation of LBP based on actual shape of mass according to sparse matrix has been shown in Figure
Lookup table approach for LBP computation from shape of mass in ROI.
Process for computation of LBP descriptor from shape of mass in ROI. (a) Original image, (b) 3 × 3 window for selection of foreground pixels, (c) lookup table, (d) decision making process, (e) LBP computation from selected foreground pixels.
Input: Output: LBP features Initialize: Radius Mask = [1 2 4 8 16 32 64 128 0] Sliding window coordinates: Count = 1 //number of pixels in for for //prepare local circular window center_pixel = //Arrange local neighborhoods of Lookup_table ( //count number of pixels greater than zero //select pixel position from lookup-table for computation of LBP if LBP_code(count,:) = count = count + 1 end end end //compute histogram of LBP codes LBP_descriptor = LBP_descriptor/count //scale invariant
The authors [
Figure
LBP code computation using sparse curvelet subband coefficients.
In this work, we have analyzed extracted ROI from mammogram using normal-abnormal, benign-malignant, and normal-malignant classes with ANN, SVM, and KNN classifiers. The detailed description of ANN classifier has been given in [
(a) FP reduction by clusters marked on original image, (b) FP reduction by thresholding, (c) FP reduction by sparse curvelet coefficient-based LBP, and ANN.
Load input image (img1) Apply CLAHE algorithm and obtain enhanced image (img2) Process img1 and img2 and obtain enhanced image using procedure given in Algorithm Remove pectoral muscle using proposed approach (Section Extract neighbourhood features for each pixel and apply SOM clustering Obtain clustered image and separate out the tumorous cluster Extract detected regions Extract Sparse Curvelet Coefficients (Subband) up to 2 level from each ROI Extract Sparse LBP code for each subband and obtain a combined feature vector for each ROI Classify each ROI into tumorous and nontumorous class Map each TP region on original mammogram (img1) end
The proposed method has been tested and validated using three classifiers and three clinical mammographic image datasets.
The mini-MIAS [
The DDSM [
This dataset [
The remaining 55 patients were examined with “GE Medical Senograph System” (Scanner2) providing 8-bit true color mammogram image in DICOM format of 4096 × 3328 or 2294 × 1914 pixels each measuring size 50 × 50
The segmentation using SOM that detects suspicious mass regions is considered as TP whereas from nonmass is taken as FP. From Table
Result of SOM segmentation.
Dataset used | Result of SOM clustering and threshold | TPR (true-positive rate) = TP/#lesions | FPPI (false-positive per image) = FP/#images | |||
---|---|---|---|---|---|---|
Mass | Segmented nonmass (FP) | Total (#) images | ||||
Segmented (TP) | Lost | |||||
MIAS | 108 | 7 | 273 | 322 | (108/115) = 0.94 | (273/322) = 0.85 |
DDSM | 140 | 10 | 1203 | 250 | (140/150) = 0.93 | (1203/250) = 4.81 |
TMCH | 172 | 8 | 837 | 360 | (172/180) = 0.95 | (837/360) = 2.32 |
From extracted ROIs, the minimum patch size is 25 × 22 pixels whereas the maximum size is 1152 × 1356 pixels. Tables
Reduction in curvelet coefficients for sample mammograms from MIAS and DDSM dataset.
Sr. No. | MIAS | DDSM | ||||||
---|---|---|---|---|---|---|---|---|
ROI Size | Total number of curvelet coefficients from subbands | Total number of selected curvelet coefficients from subbands | % reduction in curvelet coefficients | ROI Size | Total number of curvelet coefficients from subbands | Total number of selected curvelet coefficients from subbands | % reduction in curvelet coefficients | |
1 | 124 × 138 | 1,03,911 | 77,133 | 25.77 | 192 × 187 | 2,16,729 | 1,68,333 | 22.33 |
2 | 179 × 138 | 1,50,123 | 1,26,142 | 15.97 | 294 × 291 | 5,18,267 | 3,66,680 | 29.25 |
3 | 51 × 116 | 36,815 | 33,421 | 9.22 | 145 × 207 | 1,81,663 | 1,48,765 | 18.11 |
4 | 83 × 83 | 42,449 | 36,653 | 13.65 | 169 × 168 | 1,71,873 | 1,54,752 | 9.96 |
5 | 84 × 76 | 39,115 | 35,815 | 8.44 | 182 × 248 | 2,72,517 | 2,19,822 | 19.34 |
6 | 74 × 83 | 37,767 | 34,448 | 8.79 | 213 × 349 | 4,49,783 | 3,01,359 | 33.00 |
7 | 53 × 64 | 20,969 | 18,621 | 11.20 | 578 × 412 | 14,33,195 | 6,62,072 | 53.80 |
8 | 70 × 44 | 18,899 | 16,610 | 12.11 | 215 × 219 | 2,86,429 | 2,42,935 | 15.18 |
9 | 80 × 66 | 32,409 | 29,454 | 9.12 | 420 × 428 | 10,82,461 | 5,01,209 | 53.70 |
10 | 69 × 86 | 36,783 | 33,552 | 8.78 | 203 × 307 | 3,75,871 | 2,66,829 | 29.01 |
11 | 59 × 116 | 42,019 | 38,442 | 8.51 | 226 × 262 | 3,57,209 | 2,76,763 | 22.52 |
12 | 81 × 101 | 50,637 | 46,122 | 8.92 | 159 × 194 | 1,87,563 | 1,48,741 | 20.70 |
13 | 41 × 84 | 21,427 | 18,907 | 11.76 | 718 × 686 | 29,61,127 | 7,62,561 | 74.25 |
14 | 69 × 141 | 60,641 | 52,427 | 13.54 | 409 × 550 | 13,52,439 | 9,04,829 | 33.10 |
15 | 60 × 62 | 22,925 | 20,475 | 10.69 | 524 × 375 | 11,84,671 | 7,66,522 | 35.30 |
16 | 96 × 101 | 59,647 | 54,235 | 9.07 | 311 × 275 | 5,17,433 | 3,97,737 | 23.13 |
17 | 136 × 139 | 1,13,727 | 66,352 | 41.56 | 319 × 320 | 6,14,129 | 4,12,606 | 32.81 |
18 | 55 × 94 | 31,359 | 28,305 | 9.74 | 313 × 447 | 8,42,855 | 5,10,482 | 39.43 |
19 | 157 × 140 | 1,32,373 | 1,07,000 | 19.17 | 291 × 517 | 9,07,903 | 6,37,412 | 29.79 |
20 | 156 × 130 | 1,23,007 | 90,834 | 26.15 | 370 × 837 | 18,64,889 | 8,98,236 | 51.83 |
Average | 58,850 | 48,247 | 14 | Average | 7,88,950 | 4,37,432 | 32 |
Reduction in curvelet coefficients for sample mammograms from TMCH Scanner1 and Scanner2 dataset.
Sr. no. | TMCH: Scanner 1: “GE Medical Senograph System” | TMCH: Scanner 2: “Hologic Selenia System” | ||||||
---|---|---|---|---|---|---|---|---|
ROI size | Total number of curvelet coefficients from subbands | Total number of selected curvelet coefficients from subbands | % reduction in curvelet coefficients | ROI size | Total number of curvelet coefficients from subbands | Total number of selected curvelet coefficients from subbands | % reduction in curvelet coefficients | |
1 | 459 × 412 | 11,40,617 | 7,94,885 | 30.31 | 291 × 278 | 4,89,243 | 3,17,014 | 35.20 |
2 | 548 × 513 | 16,93,403 | 11,17,873 | 33.99 | 545 × 246 | 8,09,627 | 5,31,604 | 34.34 |
3 | 415 × 303 | 7,58,323 | 4,45,585 | 41.24 | 560 × 483 | 16,30,583 | 10,68,439 | 34.47 |
4 | 645 × 495 | 19,51,443 | 11,45,580 | 41.29 | 782 × 510 | 24,00,137 | 12,75,073 | 46.87 |
5 | 437 × 691 | 18,16,651 | 9,85,120 | 45.77 | 87 × 141 | 75,773 | 67,871 | 10.43 |
6 | 812 × 500 | 24,39,937 | 12,87,065 | 47.25 | 311 × 185 | 3,48,565 | 2,40,546 | 30.99 |
7 | 468 × 379 | 10,66,333 | 7,10,242 | 33.40 | 262 × 348 | 5,50,303 | 2,82,821 | 48.61 |
8 | 673 × 582 | 23,55,589 | 17,30,915 | 26.52 | 610 × 440 | 16,14,515 | 8,42,876 | 47.79 |
9 | 250 × 201 | 3,04,513 | 2,35,670 | 22.61 | 949 × 391 | 22,27,209 | 13,59,338 | 38.97 |
10 | 525 × 488 | 15,44,691 | 11,61,942 | 24.78 | 365 × 385 | 8,46,523 | 6,04,473 | 28.59 |
11 | 488 × 779 | 22,87,547 | 16,11,385 | 29.56 | 393 × 247 | 5,85,063 | 4,90,474 | 16.17 |
12 | 1434 × 966 | 83,26,581 | 37,42,277 | 55.06 | 341 × 301 | 6,18,111 | 4,10,542 | 33.58 |
13 | 348 × 421 | 8,81,701 | 6,16,501 | 30.08 | 523 × 702 | 22,06,097 | 8,00,057 | 63.73 |
14 | 460 × 530 | 14,67,227 | 7,99,885 | 45.48 | 370 × 284 | 6,32,539 | 4,23,727 | 33.01 |
15 | 398 × 450 | 10,78,441 | 8,28,064 | 23.22 | 344 × 202 | 4,18,955 | 3,02,427 | 27.81 |
16 | 247 × 272 | 4,04,401 | 3,44,822 | 14.73 | 264 × 188 | 2,99,997 | 2,31,926 | 22.69 |
17 | 411 × 305 | 7,57,657 | 4,05,919 | 46.42 | 233 × 247 | 3,46,983 | 2,67,686 | 22.85 |
18 | 286 × 344 | 5,93,155 | 4,48,566 | 24.38 | 370 × 291 | 6,50,701 | 4,86,125 | 25.29 |
19 | 417 × 207 | 5,23,477 | 4,29,755 | 17.90 | 680 × 483 | 19,79,543 | 10,52,793 | 46.82 |
20 | 463 × 458 | 12,75,021 | 8,57,608 | 32.74 | 202 × 266 | 3,24,295 | 2,15,042 | 33.69 |
Average | 16,33,335 | 9,84,983 | 33 | Average | 9,52,738 | 5,63,543 | 34 |
From Figures
Average classification rate for TMCH dataset.
Average classification rate for MIAS and DDSM dataset.
Average classification rate for MIAS and DDSM dataset.
Average classification rate for MIAS and DDSM dataset.
Data augmentation has been used for some classes to maintain balance between two classes, to improve performance, and to learn more powerful model. Table
Number of ROIs resulted in FP reduction using curvelet-based LBP (without sparse) & ANN classification at training and validation stage.
Class | Dataset used | Benign/malignant mass | Nonmass/benign mass | Total (#) images | TPR (true-positive rate) = TP/#lesions | FPPI (false-positive per image) = FP/# images | ||||
---|---|---|---|---|---|---|---|---|---|---|
Previous stage | Selected (TP) | Lost (FN) | Previous stage | Selected (TN) | Lost (FP) | |||||
Normal vs abnormal | MIAS | 108 ∗ 2 = 216 | 203 | 13 | 273 | 257 | 16 | 315 | (203/216) = 0.94 | (16/315) = 0.05 |
DDSM | 140 ∗ 4 = 560 | 465 | 95 | 1203 | 1095 | 108 | 240 | (465/560) = 0.83 | (108/240) = 0.45 | |
|
||||||||||
Benign vs malignant | MIAS | 49 | 49 | 0 | 59 | 57 | 2 | 108 | (49/49) = 1.00 | (2/108) = 0.02 |
DDSM | 46 ∗ 2 = 92 | 91 | 1 | 94 | 91 | 3 | 140 | (91/92) = 0.99 | (3/140) = 0.02 | |
|
||||||||||
Normal vs malignant | MIAS | 49 ∗ 4 = 196 | 184 | 12 | 273 | 254 | 19 | 256 | (184/196) = 0.94 | (19/256) = 0.07 |
DDSM | 46 ∗ 4 = 184 | 180 | 4 | 1203 | 1143 | 60 | 146 | (180/184) = 0.98 | (60/146) = 0.41 | |
TMCH: Scanner1 | 107 ∗ 4 = 428 | 416 | 12 | 605 | 551 | 54 | 217 | (416/428) = 0.97 | (54/217) = 0.25 | |
TMCH: Scanner2 | 65 ∗ 4 = 260 | 255 | 5 | 232 | 214 | 18 | 135 | (255/260) = 0.98 | (18/135) = 0.13 |
∗Augmentation of image.
Similarly, Table
Number of ROIs resulted in FP reduction using sparse curvelet coefficient-based LBP & ANN classification at training and validation stage.
Class | Dataset used | Benign/malignant mass | Nonmass/benign mass | Total (#) images | TPR (true-positive rate) = TP/#lesions | FPPI (false-positive per image) = FP/# images | ||||
---|---|---|---|---|---|---|---|---|---|---|
Previous stage | Selected (TP) | Lost (FN) | Previous stage | Selected (TN) | Lost (FP) | |||||
Normal vs abnormal | MIAS | 108 ∗ 2 = 216 | 201 | 15 | 273 | 265 | 8 | 315 | (201/216) = 0.93 | (8/315) = 0.02 |
DDSM | 140 ∗ 4 = 560 | 516 | 44 | 1203 | 1155 | 48 | 240 | (516/560) = 0.92 | (48/240) = 0.2 | |
|
||||||||||
Benign vs malignant | MIAS | 49 | 48 | 1 | 59 | 59 | 1 | 108 | (48/49) = 0.98 | (1/108) = 0.01 |
DDSM | 46 ∗ 2 = 92 | 89 | 3 | 94 | 89 | 5 | 140 | (89/92) = 0.97 | (5/140) = 0.03 | |
Normal vs malignant | MIAS | 49 ∗ 4 = 196 | 192 | 4 | 273 | 259 | 14 | 256 | (192/196) = 0.98 | (14/256) = 0.05 |
DDSM | 46 ∗ 4 = 184 | 182 | 2 | 1203 | 1167 | 36 | 146 | (182/184) = 0.99 | (36/146) = 0.25 | |
TMCH: Scanner1 | 107 ∗ 4 = 428 | 424 | 4 | 605 | 593 | 12 | 217 | (424/428) = 0.99 | (12/217) = 0.05 | |
TMCH: Scanner2 | 65 ∗ 4 = 260 | 260 | 0 | 232 | 232 | 0 | 135 | (260/260) = 1.00 | (0/135) = 0 |
∗Augmentation of image.
Performance evaluation of curvelet-based LBP descriptor algorithm.
Dataset | Classification | Normal-malignant | Normal-abnormal | Benign-malignant | ||||||
---|---|---|---|---|---|---|---|---|---|---|
Classifier | Sensitivity | Specificity | AUC | Sensitivity | Specificity | AUC | Sensitivity | Specificity | AUC | |
MIAS | ANN | 0.94 | 0.93 | 0.94 | 0.94 | 0.94 | 0.94 | 1.00 | 0.97 | 0.99 |
SVM | 0.85 | 0.85 | 0.85 | 0.83 | 0.86 | 0.85 | 0.88 | 0.84 | 0.86 | |
KNN | 0.67 | 0.63 | 0.65 | 0.58 | 0.57 | 0.58 | 0.62 | 0.68 | 0.63 | |
|
||||||||||
DDSM | ANN | 0.98 | 0.95 | 0.95 | 0.83 | 0.91 | 0.85 | 0.99 | 0.97 | 0.98 |
SVM | 0.97 | 0.88 | 0.92 | 0.71 | 0.91 | 0.83 | 0.94 | 0.89 | 0.92 | |
KNN | 0.96 | 0.64 | 0.87 | 0.67 | 0.90 | 0.80 | 0.87 | 0.73 | 0.79 | |
|
||||||||||
TMCH: Scanner1 | ANN | 0.97 | 0.91 | 0.94 | — | — | — | — | — | — |
SVM | 0.96 | 0.91 | 0.94 | — | — | — | — | — | — | |
KNN | 0.98 | 0.82 | 0.89 | — | — | — | — | — | — | |
|
||||||||||
TMCH: Scanner2 | ANN | 0.98 | 0.92 | 0.96 | — | — | — | — | — | — |
SVM | 0.97 | 0.90 | 0.94 | — | — | — | — | — | — | |
KNN | 0.92 | 0.83 | 0.88 | — | — | — | — | — | — |
Performance evaluation of proposed algorithm.
Dataset | Classification | Normal-malignant | Normal-abnormal | Benign-malignant | ||||||
---|---|---|---|---|---|---|---|---|---|---|
Classifier | Sensitivity | Specificity | AUC | Sensitivity | Specificity | AUC | Sensitivity | Specificity | AUC | |
MIAS | ANN | 0.98 | 0.95 | 0.96 | 0.93 | 0.97 | 0.95 | 0.97 | 1.00 | 0.98 |
SVM | 0.88 | 0.83 | 0.85 | 0.85 | 0.82 | 0.84 | 0.84 | 0.92 | 0.87 | |
KNN | 0.55 | 0.51 | 0.53 | 0.55 | 0.63 | 0.56 | 0.61 | 0.67 | 0.61 | |
|
||||||||||
DDSM | ANN | 0.99 | 0.97 | 0.98 | 0.92 | 0.96 | 0.93 | 0.97 | 0.95 | 0.96 |
SVM | 0.99 | 0.92 | 0.96 | 0.89 | 0.96 | 0.92 | 0.94 | 0.92 | 0.93 | |
KNN | 0.98 | 0.73 | 0.92 | 0.74 | 0.90 | 0.82 | 0.89 | 0.77 | 0.83 | |
|
||||||||||
TMCH: Scanner1 | ANN | 0.99 | 0.98 | 0.98 | — | — | — | — | — | — |
SVM | 0.98 | 0.96 | 0.97 | — | — | — | — | — | — | |
KNN | 0.99 | 0.92 | 0.95 | — | — | — | — | — | — | |
|
||||||||||
TMCH: Scanner2 | ANN | 1.00 | 1.00 | 1.00 | — | — | — | — | — | — |
SVM | 1.00 | 0.98 | 0.99 | — | — | — | — | — | — | |
KNN | 0.96 | 0.92 | 0.94 | — | — | — | — | — | — |
However, from Table
Representation of fully automatic CAD system for breast cancer using (a) sample mammograms from MIAS, DDSM, and TMCH datasets, (b) preprocessed mammograms, (c) clustered image, (d) TP and FP marked on mammogram, (e) TP marked by thresholding, (f) TP marked by using LBP descriptor based on sparse curvelet coefficients.
Table
Comparison of classification accuracy, AUC, and FP/image values from different approaches in breast cancer diagnosis.
Author | Database | Method | Classifier | Result | AUC | FP/image |
---|---|---|---|---|---|---|
Eltoukhy et al. [ |
MIAS | Biggest curvelet coefficients as a feature vector | Euclidean classifier | 94.07% | — | — |
Eltoukhy et al. [ |
98.59 | — | — | |||
Eltoukhy et al. [ |
SVM | 97.3 | — | — | ||
Dhahbi et al. [ |
Mini-MIAS | Curvelet moments | KNN | 91.27 | — | — |
DDSM | 86.46 | — | — | |||
Bruno et al. [ |
DDSM | Curvelet + LBP | SVM | 85 | 0.85 | — |
PL | 94 | 0.94 | — | |||
da Rocha et al. [ |
DDSM | LBP | SVM | 88.31 | 0.88 | — |
Kanadam and Chereddy [ |
MIAS | Sparse ROI | SVM | 97.42 | — | — |
Pereira et al. [ |
DDSM | Wavelet and Wiener filter | Multiple thresholding, wavelet, and GA | — | — | 1.37 |
Liu and Zeng [ |
DDSM, FFDM | GLCM, CLBP, and geometric features | SVM | — | — | 1.48 |
De Sampaio et al. [ |
DDSM | LBP | DBSCAN | 98.26 | 0.19 | |
Zyout et al. [ |
DDSM | Second order statistics of wavelet coefficients (SOSWC) | SVM | 96.8 | 0.97 | 0.018 |
MIAS | 95.2 | 96.6 | 0.029 | |||
Casti et al. [ |
DDSM | Differential features | Fisher linear discriminant analysis (FLDA) | — | — | 1.68 |
MIAS | 2.12 | |||||
FFDM | 0.82 | |||||
Proposed method | MIAS | LBP based on sparse curvelet subband coefficients | ANN | 98.57 | 0.98 | 0.01 |
DDSM | 98.70 | 0.98 | 0.03 | |||
TMCH: Scanner1 | 98.30 | 0.98 | 0.05 | |||
TMCH: Scanner2 | 100 | 1 | 0 |
A fully automatic CAD system, which can accurately locate the tumor on a mammogram and reduces FPs, has been proposed. The developed CAD system consists of preprocessing, SOM clustering, ROI extraction, sparse LBP feature computation based on sparse Curvelet coefficients, and finally, FP reduction using ANN classifier.
The proposed algorithm presents a novel concept of extraction of curvelet coefficients according to irregular shape of mass is called as sparse curvelet coefficients and computation of LBP. The analysis proves that the FPs are reduced significantly from 0.85 to 0.01 FP/image for MIAS, 4.81 to 0.03 FP/image for DDSM and 2.32 to 0.00 FP/image for TMCH. The ANN classifier showed best results as AUC = 0.98 and accuracy = 98.57% for MIAS in benign-malignant classification, AUC = 0.98 and accuracy = 98.70% for DDSM in normal-malignant classification, AUC = 0.98 and accuracy = 98.30% for TMCH: Scanner1, and AUC = 1 and accuracy = 100% for TMCH: Scanner2 in normal-malignant classification as compared with SVM and KNN classifier. The performance of LBP features and LBP features based on sparse curvelet coefficients are nearly same which show that the proposed algorithm is suitable for cancer breast tissue diagnosis.
In future, the reduced curvelet coefficients can be used to extract local ternary patterns and other local descriptor and local directional patterns, etc. The present work deals with mammogram with single mass; this can be further extended for multiple mass models with multiple LBP features based on sparse curvelet coefficients.
In this research, we have used two publicly available datasets MIAS and DDSM. These datasets can be found here in [
The authors declare that they have no conflicts of interest.
The TMCH database for this work was given by Department of Radiodiagnosis, Tata Memorial Cancer Hospital, Mumbai.