Performance Comparison of Pre-trained Residual Networks for Classification of the Whole Mammograms with Smaller Dataset

• Published in: 2020 IEEE Student Conference on Research and Development (SCOReD) pp. 368 - 373
• Main Authors: Bagchi, Susama, Mohd, Mohd Norzali Haji, Debnath, Sanjoy Kumar, Nafea, Marwan, Suriani, Nor Surayahani, Nizam, Yoosuf
• Format: Conference Proceeding
• Language: English
• Published: IEEE 27.09.2020
• Subjects: Breast cancer; Computer-aided Detection (CAD) System; Convolutional neural networks; Deep Convolutional Neural Network (DCNN); Digital Mammogram Classification; Mammography; Residual Network (ResNet); Residual neural networks; Sensitivity and specificity; Testing; Training
• ISSN: 2643-2447
• DOI: 10.1109/SCOReD50371.2020.9250939

The false-positive breast cancer cases detected by radiologists and Computer-aided Detection (CAD) systems increase the medical cost and patient discomfort due to the unnecessary breast biopsies. These available CAD systems were developed using traditional machine learning techniques for breast cancer diagnosis. A noteworthy progress is happening in cancer diagnosis after the introduction of deep learning in Convolutional Neural Networks (CNNs) for CAD development. This paper compares the performance of three pre-trained Residual Networks (ResNets), i.e., ResNet18, ResNet50, and ResNet101 with increased image input layer size of 512\times 512\times 3 for the classification of the pre-processed whole mammograms into normal, benign, and malignant categories. INbreast dataset was pre-processed and then these pre-processed whole breast images were segregated into three categories based on the ground truths. Original and modified networks were developed by replacing the last three layers of the selected ResNets to match the output category along with the image input layer. Data augmentation and transfer learning were applied to overcome the overfitting issue due to smaller dataset. The developed models were tested and the attained training and testing accuracies, sensitivity, and specificity were compared to evaluate their performances. It was observed that ResNet50 with an image input layer of size 512\times 512\times 3 provided best results after five-fold training and the test accuracy was 79.27% with the average sensitivity and specificity of 0.76, and 0.89, respectively for three categories. This experimental work is significant as it proves that the increased image input layer size has a considerable effect in classifying the whole mammograms. Further development will be done with a balanced dataset and other pre-trained deep networks will also be tried.