Classification: lesion and disease recognition

In this chapter, we introduce how to design deep neural networks for medical image classification. We begin by introducing several design principles, including 1) the choice of deep neural networks, 2) the choice of classification tasks and objectives, 3) what is and when to apply transfer learning, and 4) what is and when to apply multitask learning. Next, we provide two case studies and show how these design principles are applied to address the skin lesion and disease recognition problems. Specifically, we introduce two classification scenarios, multiclass and multilabel classification. We investigate if one classification scenario could be better than the other when it comes to skin lesion and disease recognition. Then, we position the skin lesion recognition problem under a multitask learning scenario and investigate the benefit of leveraging additional tasks for skin lesion recognition.