Analysis and Classification of Breast Abnormalities Using Ultrasound Images

Abstract

The present research work has been carried out with an aim to enhance the diagnostic potential of B-mode ultrasound imaging modality for the diagnosis of breast abnormalities. To achieve this objective exhaustive experiments have been carried out in the present research work to (a) analyse the effect of despeckle filtering algorithms on breast ultrasound images, (b) analyse the effect of despeckle filtering algorithms on segmentation of breast tumors, (c) analyse the effect of despeckle filtering algorithms on classification of breast tumors, (d) design an efficient local binary pattern (LBP) based CAD system for classification of breast tumors, (e) design an efficient convolutional neural network based CAD system for classification of breast tumors. For carrying out the experiments a comprehensive dataset of 100 B-mode breast ultrasound images comprising of cysts, fibroadenomas, lipomas in benign category, ductal and lobular carcinomas in malignant category has been taken from a standard benchmark database, ultrasoundcases.info. Initially exhaustive experimentations have been carried out to analyze the effect of 42 despeckle filtering algorithms taken from various filter categories namely (a) Local statistics based filters, (b) Fourier filters, (c) Fuzzy filters, (d) Multiscale filters, (e) Non-local mean filters, (f) Non-linear iterative filters, (g) Total variation filters and (h) Hybrid filters. The resultant despeckled images have been used for objective assessment and subjective assessment. For the objective assessment, an image quality metric named structure and edge preservation index (SEPI) has been proposed. This index quantifies the edge preservation and structure preservation capability of the filtering algorithm. Based on the results of the objective evaluation, out of 42 filters, 06 best performing despeckle filtering algorithms namely Lee sigma, BayesShrink, Detail preserving anisotropic diffusion (DPAD), Fourier ideal (FI), Fourier Butterworth (FB), and Homomorphic Fourier Butterworth (HFB) filters have been selected that result in controlled despeckling of the images. The images despeckled by these 06 best performing despeckle filtering algorithms have further been used for the subjective assessment based on the radiologist’s grading. On the basis of both objective and subjective assessment, DPAD filter has been selected as an optimal filter for pre-processing the breast ultrasound images. The 100 original breast ultrasound images and 100 despeckled images pre-processed using 06 best performing despeckle filtering algorithms have been considered for analysing the effect of despeckle filtering algorithms on segmentation of breast tumors. Effect of despeckle filtering algorithms on segmentation of breast tumors- The original images as well as the images despeckled by the 06 best performing despeckling filtering algorithms have been subjected to the Chan and Vese active contour method for segmentation of breast tumors. The objective assessment of the segmentation algorithm has been carried out by computing the Jaccard index and subjective assessment has been carried out by the experienced participating radiologist. Based on the results of the study the DPAD has been selected as an optimal despeckle filtering algorithm resulting in efficient segmentation of breast tumors. Exhaustive experiments were carried out for analysing the effect of best performing despeckle filtering algorithms on classification of breast tumors- Initially four PCA-SVM based CAD systems based on (a) texture and morphological features computed from original images (b) texture and morphological features computed from despeckled images, (c) texture features computed from original images and morphological features computed from despeckled images and (d) texture features computed from despeckled images and morphological features computed from original images. It was observed that the PCA-SVM based CAD system design based on texture features computed from original images and morphological features computed from images despeckled using DPAD filter yielded highest classification accuracy of 96.0 % with individual class accuracy values of 95.2 % and 96.6 % respectively. Four LBP based CAD system designs were implemented using classifiers namely, PCA-SVM, ANFC-LH, GA-SVM and SAE-SM by computing LBP based texture features computed from original images and morphological features computed from images despeckled using DPAD filter- It was observed that the LBP based CAD system with ANFC-LH classifier yielded the highest accuracy of 96.0 % with individual class accuracy values of 90.4 % and 100 % for benign and malignant classes, respectively. Further to compare the performance of the existing conventional approaches with the state of the art deep learning techniques- Initially four different transfer learning based CNN models with GoogleNet, VGG-19, ResNet-18 and SqueezeNet architecture have been implemented. It was observed that CNN based model based on GoogLeNet architecture yielded highest accuracy. Accordingly an optimal CNN based CAD system design based on deep feature extraction by GoogLeNet architecture and classification by ANFC-LH was designed yielding highest classification accuracy of 98.0 % with ICA values of 100 % and 96.6 % for benign and malignant cases respectively.