A Multi-Task Learning Based Approach for Efficient Breast Cancer Detection and Classification

Abstract

Automatic segmentation and classification of breast tumours in ultrasound images using deep learning approaches can help early detect breast cancer. Such predictive modelling can potentially significantly improve the survival chances of the involved patients. Most of the typical deep convolutional neural network (CNN) based approaches consider segmentation and classification tasks separately. But this loses important supervisory information to help achieve better model training. This work proposes the integrated learning of both of these tasks in an end-to-end manner, using a multi-task learning based approach. More specifically, a convolutional encoder-decoder based architecture is coupled with a residual CNN for performing segmentation and classification together. The level-wise feature maps from both the encoder and decoder parts of the segmentation network are utilized for classification in the proposed approach. From experimental analysis on a publicly available breast ultrasound image (BUSI) dataset, it has been observed that the proposed approach can achieve impressive performances, both with respect to tumour segmentation and classification. A mean test set AUC of 0.97 and a mean dice score of 0.74 is achieved, establishing a new state-of-the-art performance on the BUSI dataset. From the impressive experimental observations, it can be concluded that learning to perform both segmentation and classification simultaneously can have a very high positive impact on the overall quality of the predictive model. Such observations suggest that the proposed approach can be beneficial in providing real-time decision support to the involved diagnostic radiologists, which can help improve the survival chances of the corresponding patients.