Adaptive Self-Supervised Learning of Morphological Landscape for Leukocytes Classification in Peripheral Blood Smears

Diagnosis of hematological disorders relies on cytomorphology and abundance of white blood cells (WBC) in peripheral blood smear (PBS). Hematology analyzers offer automated cell classification but cannot supersede manual review, especially in cancer patients where disease or treatment-induced morphological shifts necessitate frequent label corrections. To overcome both the rigid cell type definition and inefficient label usage, we developed a self-supervised model trained on image triplets with a lightweight EfficientNetV2-B0 encoder. With the learned morphological landscape, cell-type labels can be obtained using simple classifiers with tunable support sets, achieving a testing 9-way classification <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\boldsymbol{F}_{\mathbf{1}}$</tex> score of 96.2%. Moreover, our model readily generalizes to different label sets as demonstrated by predicting 11 morphological attributes. Active learning was used to further increase label efficiency without sacrificing performance. To enable wider adoption, the model was made accessible as a web application, HemoSight. We anticipate adaptive, accurate, and efficient self-supervised image classification to accelerate clinical workflow with morphological insight.