Diagnosis of Acute Lymphoblastic Leukemia through White Blood Cell Segmentation from Multi-Microscope Imaging using Deep Learning

Abstract

Acute Lymphoblastic Leukemia, as a severe type of blood cancer, can be characterized by the uncontrolled increase of white blood cells, and it is crucial to make an early and accurate diagnosis to provide effective treatment. The traditional method of using manual examination of peripheral blood smear slides is slow and very reliant on the expertise of the examiner, which introduces variability. To address this, this research introduces an interpretable deep learning method that integrates both the segmentation and classification of WBCs based on the images created by multi-microscopic scanner imaging to better reflect a real-world lab setting. A combined dataset was created using publicly available blood smear images and a custom LeukemiAttri dataset. Training of the segmentation model was done through manual annotation of preprocessed images. A refined Segment Anything Model was used as a robust method of WBC segmentation, achieving an average segmentation accuracy of 87%. The segmented cells have been subsequently grouped as healthy or leukemic on the basis of five deep learning models, namely DenseNet-169, DenseNet-121, VGG16, VGG19, and MobileNet, with DenseNet-169 reporting the best classification accuracy of 92%. Compared to previous single-scanner approaches, this multi-scanner framework achieved better generalization and higher accuracy, offering a practical solution for reliable automated leukemia detection in clinical practice.