Feature Fusion for Segmentation and Classification of Skin Lesions

Abstract

Automated segmentation and classification of dermoscopy images are two crucial tasks for early detection of skin cancers. Deep models trained for individual task ignore the relationship of the two tasks and lack the diagnostic proposals or explanation for diagnosis results in practice. We assume that features extracted with segmentation models and classification models on the same dataset are highly related and the two tasks have potentials to boost each other when trained together. In this paper, we propose a combined-learning network (CLNet) consisting of a classification network, a segmentation network and a feature fusion module for segmentation and classification of skin lesions. Particularly, the feature fusion module fuses features extracted by the classification branch and segmentation branch and outputs fused features for the two tasks respectively. In this way, the information shared by the two branches can be fully exploited and the performance of two tasks can be mutually improved. Experimental results demonstrate that the proposed model can achieve promising performance on the public skin disease dataset.