Acute lymphoblastic leukemia diagnosis and aubtype segmentation in blood smears using CNN and U-Net

Abstract

Acute lymphoblastic leukaemia (ALL) is a severe disease requiring invasive, expensive, and time-consuming diagnostic tests for definitive diagnosis. Initial diagnosis using blood smear pictures (BSP) is crucial but challenging due to the similar indications and symptoms of ALL, often leading to misdiagnoses. This study presents a custom approach using Convolutional Neural Networks (CNNs) to detect all cases and categorize subtypes. Utilizing publicly available databases, the study includes 3562 blood smear images from 89 patients. The innovative combination of U-Net for segmentation and various CNN architectures (U-Net, MobileNetV2, InceptionV3, ResNet50, NASNet) for feature extraction, with DenseNet201 being the most effective, forms the core of this method. The U-Net model achieved a segmentation accuracy of 98% by recognizing patterns within blood smear images. Following segmentation, CNN architectures extracted high-level features, with DenseNet201 proving the most effective in diagnostic and classification tasks. Our proposed custom CNN model achieved a test accuracy of 98%, with a training accuracy of 99.31% and validation accuracy of 97.09%. This approach enables an accurate distinction between ALL and non-pathologic cases.