Leukemic B-Lymphoblast Cell Detection with Monte Carlo Dropout Ensemble Models

Abstract

Automatic detection of Leukemic B-lymphoblast cells from normal cells is both crucial and difficult with subject variability being a major obstacle. While deep learning models are widely adopted in image recognition and demonstrate state-of-the-art performance, deep learning models in their vanilla form do not provide an estimate of prediction uncertainty. This is especially problematic when there is a high discrepancy between training and testing data such as in blood smear images where subject level variability and sample noise from staining and illumination are non-negligible. To address these issues, we propose a novel ensemble method by weighting each base model according to its respective predictive confidence obtained from Monte Carlo dropout. Our ensemble method can effectively prevent overfitting and increases classification performance compared to any single model result. Furthermore, we demonstrate the model’s ability to identify samples that do not fall in the training data distribution. We achieved 0.8929 prediction score according to the leaderboard on the SBILab challenge website.