Early Computer-Aided Diagnose in Medical Environments: A Deep Learning Based Lightweight Solution

Abstract

The use of artificial intelligence in healthcare systems has helped doctors to automate many tasks and has avoided unnecessary patient hospitalizations. Mobile devices have grown in computing capacity and enhanced image acquisition capabilities, which enable the implementation of more powerful outpatient services. In this chapter we propose a lightweight solution to the diagnose of skin lesions. Specifically, we focus on the melanoma classification whose early diagnosis is crucial to increase the chances of its cure. Computer vision algorithms can be used to analyze dermoscopic images of skin lesions and decide if these correspond to benign or malignant tumors. We propose a deep learning solution by means of the adaptation of the attention residual learning designed for ResNets to the EfficientNet networks which are suitable for mobile devices. A comparison is made of this mechanism with other attention mechanisms that these networks already have incorporated. We maintain the efficiency of these networks since only one extra parameter per stage needs to be trained. We also test several pre-processing methods that perform color corrections of skin images and sharpens its details improving the final performance. The proposed methodology can be extended to the early detection and recognition of other forms of skin lesions.