Breast Cancer Diagnosis Using Lightweight Deep Convolution Neural Network Model

Abstract

In the past few decades, breast cancer has rapidly increased the death rate among women worldwide. An early diagnosis of such fatal disease is important for the best treatment and death rate reduction. Automatic diagnosis of breast cancer from histopathological images using artificial intelligence (AI) based methods is a top-priority research area in the biomedical field. However, automatic detection is challenging due to high resolution of histopathology images and the tremendous amount of parameters required by deep AI models. Due to high computational complexity and bulky memory usage, deep models suffer from inefficient inference that limits their application in resource-constrained platforms. To address this problem, a fast cancer detection strategy has been proposed to overcome the computational cost issue of deep automatic systems. Instead of directly using input images the wavelet transform (WT) is applied to decompose the images into different frequency bands and then only low frequency bands are subjected to our proposed lightweight deep convolutional neural network (CNN). The lightweight deep model is designed using invertible residual block module. The incorporation of invertible residual block module in the deep CNN model and the use of WT considerably reduces the computational cost of the proposed model, without a noticeable accuracy downgrade Further, the effect of various machine vision classifiers i.e. support vector machine (SVM), softmax, and K nearest neighbor classifier (KNN) on model performance is analyzed. Experiments are performed using three publicly available benchmark histopathology image datasets. The proposed model has shown multi-class classification accuracy of 96.25%, and 99.8% and 72.2%, on the international conference on image analysis and recognition (ICIAR 2018), BreakHis and Bracs datasets, respectively. The reported inference time per image of the proposed model is 0.67s, and 0.21s for ICIAR 2018 and BreakHis and Bracs images, respectively.