Deep Learning for Biology

Chapter 5. Detecting Skin Cancer in Medical Images

In previous chapters, we focused on small-scale biological phenomena, such as the molecular properties of proteins, DNA sequences, and drug molecules. In this chapter, we will zoom out to a larger biological scale, applying deep learning to analyze tissue-level and disease-related processes. Specifically, we will train a skin cancer detection model to classify images of skin into various cancerous or benign categories.

This is an exciting application because deep learning models have made significant strides in skin analysis, with studies achieving dermatologist-level accuracy in distinguishing malignant from benign lesions since at least 2018.¹ While challenges remain in integrating these models into clinical workflows—such as regulatory approval, data standardization, and prediction explainability—their potential to assist medical professionals by enhancing early detection and reducing unnecessary biopsies is highly promising.

We will be using skin cancer image data from the International Skin Imaging Collaboration (ISIC), a project dedicated to advancing skin cancer imaging research and providing standardized datasets. Over the years, ISIC has released a range of challenges focused on skin lesion classification and pathology, with an increasing number of images available. To learn more, you can read this review paper of ISIC datasets and benchmarks.²

The dataset we will use is available as the “Skin Cancer ISIC” challenge ...