The problem
Melanoma decision support requires accounting for lesion diversity, class imbalance and model limitations. DIAMELEX is an ANR-funded project focused on diagnosis by example.
My contribution
I contributed to developing and evaluating deep learning approaches for skin image analysis. This work resulted in five publications within the collaborative, ANR-funded DIAMELEX project.
Approaches explored
CNN ensembles, decoupled training for long-tailed distributions and combining neural networks with game theory. Related papers and repositories describe the methods; these are not all components of one clinical system.
Available evidence
IJMS, Sensors and DICTA publications, along with the HMLoss and CAD repositories, provide methods and code for inspection. The JAAD 2024 publication offers another connection between algorithmic research and clinical collaboration.
Limitations and lessons
A result on a research dataset does not guarantee performance in a medical centre. Evaluation conditions, study populations and code availability must be assessed paper by paper.