56 — Skin Lesion Analysis Toward Melanoma Detection: A Challenge At The 2017 International Symposium On Biomedical Imaging (ISBI), Hosted By the International Skin Imaging Collaboration (ISIC)

The International Symposium On Biomedical Imaging (ISBI) hosts an annual competition for the correct identification, segmentation, and diagnosis of skin lesions. Melanoma is the most common form of cancer in America, and is likely the easiest to detect (my words) because it manifests on the skin surface.

The International Skin Imaging Collaboration (ISIC) competition is an open contest to perform three main tasks on a large (thousands) collection of skin-lesion images:

First, participants must segment the lesion. The result of this is a binary mask that can be used to identify which pixels in the 2D scene are lesion, and which are healthy skin.

Second, participants must classify the type of lesion, based on qualities of color, texture, and distribution.

Third and finally, participants pick a most-likely diagnosis for the lesion, choosing from melanoma, nevus, and seborrheic keratosis, assigning a float probability $[0..1]$ to each.

46 final algorithms submissions competed on several success metrics in each category; the top five performers all used external data sources in addition to the ISIC data.

I found that particularly interesting; we know that data-scarcity is a dramatic shortcoming of the machine-learning world, and I think that — especially in light of spending today at JHMI Digital Health Day 2017 today — this data-scarcity problem will only get more pressing as we foray deeper into a clinical setting.