Efficient breast cancer detection via cascade deep learning network

Abstract

Breast calcifications or irregular tissue growth are major health concerns that can lead to breast cancer. To enable early management, which significantly lowers death rates, it is crucial to perform screening and determine if a tumor is benign or malignant. Building a cascade network model that bases predictions on the shape, pattern, and spread of the tumor is how this research approaches the challenge. Pre-processing of images, followed by segmentation and classification, are common methods to accomplish this. The strategy in this research employs a cascade network with UNet architecture for segmentation with a ResNet backbone for classification. To enable classification to make predictions, segmentation process involves separating tumor from the image in the form of a mask. The segmentation model's F1-score measurement came out to be 97.30%. The final decision-making layer's neural network is a straightforward 8-layer network, which follows the ResNet50 model. The proposed model's classification accuracy was 98.61%, with F1 score of 98.41%. Comparative evaluations are conducted together with the comprehensive experimental results.