br patterns incorporate a recursive process
337 patterns incorporate a recursive process to determine the
338 neighbouring pixels. However, the recursive algorithms are
339 inefficient as far as time and computational power are
340 concerned. To overcome this pitfall, a novel scheme called
341 ‘‘Sapate's neighbouring pixels selection scheme (SNPSS)’’ for
342 selecting unique pixels along eight radial directions is
344 The literature surveyed reveals that the tumour starts with
345 a tiny core and slowly starts growing in a radial pattern. This
346 type of biological tumour growth pattern has motivated us to
347 come out with SNPSS for modifying region growing method.
348 SNPSS allows us to get rid of the curse of recursive expansion
349 of region growing algorithm. The selection of neighbouring
350 pixels minimizes the repetition of the pixels which results in
351 computational effectiveness and requires no recursion at all.
Thus, segmentation using SNPSS plays a very important role in 352 delineating the exact boundary of the mass or tumour which is 353 crucial in further characterization and treatment planning 354 . The results of mammogram segmentation are depicted in 355 Fig. 2(a) and (c). 356 The segmented images with multiple candidate lesions 357 undergo one more stage to reduce the false positive lesions. 358 The proposed system filters out some of the candidate lesions 359 using few empirical rules. Empirical rules are usually used for 360 predicting some data which have no numeric expression. This 361 happens when it CCCP uncoupler is very difficult to obtain right data which 362 follows normal distribution. Empirical rules applied to any 363 random variable can give rough estimate of what your data 364 collection may look like if you were able to survey the large 365 population. In our study, these parameters include area less 366 than 2500 pixels, extent smaller than 0.15, perimeter twice that 367 of image width and eccentricity greater than 0.98. All 368 parameters and their respective values are selected under 369 the guidance of the expert radiologists involved in this study. 370 While applying the empirical rule to any of above parameters, 371 the condition that 'approximately 68% of the data falls within 372 one standard deviation of the mean' is confirmed to be true. 373 The experimental results obtained are shown in Fig. 2(b) 374 and (d). Another case with multiple lesions is shown in Fig. 3. 375 The lesions in Fig. 3(b) are paired with lesions on MLO view in 376 Fig. 3(d) using a procedure described in Section 4.3.2. Seven 377 selected geometric features concerning shape, size, margin, 378 contrast and six textural features along 3 distances, 4 379 directions based on GLCM are calculated to characterize and 380 classify the detected lesions. The features selected have 381 shown highest classification accuracy. A pool of actively 382 discriminative geometric and textural features is fed to the 383 simple SVM classifier to achieve the classification results. The 384
Please cite this article in press as: Sapate S, et al. Breast cancer diagnosis using abnormalities on ipsilateral views of digital mammograms. Biocybern Biomed Eng (2019), https://doi.org/10.1016/j.bbe.2019.04.008
Fig. 3 – FP reduction using empirical rules during the initial detection stage: (a) CC view with suspicious lesion detected; (b) multiple selected lesions after applying empirical rules for FP reduction; (c) MLO view of the same breast; (d) two lesions selected after FP reduction.
385 pixels in the background of the suspicious lesion do not play a The average of the accuracy achieved for each fold is taken as a 414 386 vital role in characterizing the mass or tumour; hence the ROI final result of the classifier. 415 387 is taken excluding the background. The textural and geometric
388 features extracted are enlisted and defined in Appendix A and 4.3. Ipsilateral-view diagnosis 416 389 Appendix B respectively.
390 Support Vector Machine (SVM) is the simple, most popular The overall design of proposed CAD system using ipsilateral 417 391 yet strong classifier for a small set of images . It focuses views is depicted in Fig. 4. The scheme generates a 418 392 more on confidence of classification to discriminate the correspondence score for each suspicious lesion from two 419 393 benign mass and malignant tumours using the respective views. The lesion is localized and confirmed as benign mass or 420 394 feature vectors as input. The decision function is defined as malignant tumour as a final result which is achieved with four 421 395 Eq. (4):
steps: (1) The suspicious lesions on both views are located 422
using arc-based method and are paired using width of arc pair,