Raluca Alexandra Jalaboi: Explaining diagnostic reasoning from a human and AI perspective
In a population of almost 8 billion people, it has been estimated that at least a third of them will experience a skin disease sometime during their lifetime. However, throughout the world countries are experiencing a scarcity of dermatologists, with waiting times for a consultation sometimes exceeding 3 months.
Automating dermatological tasks could decrease the time it takes a patient to obtain proper medical care. Convolutional neural networks (CNNs) can solve various medical tasks including skin lesion diagnosis, but one downside is the difficulty to understand the reasoning behind their predictions. Their lack of explainability decreases the trust both healthcare professionals and regulatory agencies place in products using this technology.
Various groups have been tackling the explainability problem in CNNs, but exploration is still needed to understand what diagnosis explainability means to the medical community. Dermatologists usually analyse a skin lesion by identifying its visual characteristics (basic descriptive terms, distribution, shape, topography, palpation, and additional descriptive terms), and then use these findings to infer a diagnosis. As some characteristics are irrelevant for the diagnosis, and experts disagree as to which characteristics are most important, the details of the diagnosing process are still to be understood.
This project will first attempt to explain skin lesion diagnoses from a dermatologist’s perspective by building an explainability dataset. Then, the next step will be to train CNN models to incorporate expert knowledge in their explanations. Enhanced with explainability capabilities, CNNs may solve the global issues surrounding access to trusted, quality healthcare.