Multi-task deep learning framework for leukemia cell classification and segmentation: A joint learning approach

Abstract

The classification of leukemia and segmentation of cells are tasks that have medical image analysis with the potential to enable early diagnosis and treatment. This paper propose a novel multi-task deep learning framework for leukemia cell classification and segmentation via a shared convolutional neural network backbone. We design a singular architecture that overcomes the inefficiencies of traditional single-task methods by jointly optimizing both tasks. The model proposed provides incredibly good performance with a macro-averaged F1-score of 95.09% for classification across four leukemia classes (Benign, Early, Pre, Pro). It also provides segmentation performance with IoU of 55.12% and Dice score of 67.70% on test data. The model achieves 1.87 GFLOPs indicating its computational effectiveness and high accuracy for both tasks. In tests on a comprehensive leukemia dataset, our multitask method not only saves on computing time but also enhances generalization through shared features making it suitable for real-time clinical use.