An Explainable Deep Learning Framework for Acute Lymphoblastic Leukemia Classification

Abstract

Acute Lymphoblastic Leukemia (ALL) is a type of blood cancer caused by the abnormal growth of white blood cells. Tradition-ally, doctors manually examine peripheral blood smears under a micro-scope, which can be time-consuming and varies from doctor to doctor. Deep Learning models speed up this diagnostic process. However, they do not provide any explanation for their decisions, and their reliability is also questionable. This paper presents a framework for ALL classifi-cation that provides explainability and reliability. First, the white blood cells were segmented using a U-Net model with a ResNet34 encoder. The nucleus was segmented using HSI-L*a*b color space fusion and k-means clustering. From the segmented cell and nucleus, 36 handcrafted features were extracted. An EfficientNet-B3 model was used to extract deep fea-tures from the segmented cell. Both the handcrafted and deep features were fused together and passed into a Multi-Layer Perceptron (MLP) for classification. Grad-CAM was used to provide explainability, and Monte Carlo (MC) dropout was used to measure the uncertainty in the model prediction. The proposed method achieved an accuracy of 98.85%, pre-cision of 99.29%, recall of 98.46%, specificity of 99.23%, and an AUC of 0.9981 on the ALL-IDB2 dataset. The proposed method provides high accuracy along with visual explanations and uncertainty measures.