Domain adaptation with an adversarial algorithm for blood cell classification

With Konstantins Starovoitovs we implemented an algorithm during the Helmholtz Information & Data Science Academy Data Challenge “Help A Hematologist Out”, which took place last October.

Hematogolists analyze blood smears of their patients and evaluate content of pathological blood cells that might hint at leukemia, anemia and other blood-related disorders. Images of blood smears coming from different labs vary in sharpness, brightness, contrast, scale and other properties. Therefore, one looks for an algorithm that would be agnostic towards these secondary factors and can confidently discriminate cell images regardless of their origin, which provides a use case for domain adaptation.

Blood cell images from the source datasets (left and center) and target dataset (right)

Given a dataset of labelled images of white blood cells, we applied an adversarial domain adaptation algorithm to infer labels on another unlabelled dataset, achieving scores comparable with the top submissions.

Please check Konstantins’ post on Medium