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

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