ALL-MorphNet: A Novel Hybrid Deep Learning Architecture for Automated Acute Lymphoblastic Leukemia Diagnosis in Microscopic Images

Acute lymphoblastic leukaemia (ALL) is the most common cancer in children, but early detection is difficult because its symptoms are often subtle and easy to miss. While automated diagnostic tools hold significant promise, they face numerous challenges. These include a lack of diverse data, unbalanced datasets, and the requirement for high-performance computing. To address these challenges, we developed ALL-MorphNet, a novel deep learning model designed to classify ALL cells from microscopic images. The model starts by using a convolutional stem block to capture small local details from the images. Next, a dynamic morphological block analyzes the images at multiple scales to capture both fine- and coarse-grained patterns. This step uses an attention mechanism to focus on the most important features. Subsequently, a fusion module concatenates these features without information loss. Finally, a hybrid transformer block incorporates long-range relationships into a comprehensive understanding of the image. We evaluated the model on the ALL-IDB2 dataset, achieving an accuracy of 94%, outperforming traditional CNNs, such as ResNet, VGG, and transformer-based models, which achieved 85-90% accuracy on the same dataset. Future studies will aim to improve the model practicality to a variety of datasets, multimodal inputs, and enhanced morphological processing. Additionally, we focused on optimising the framework for deployment in resource-constrained environments to ensure its applicability to real-world clinical workflows. This advancement contributes to enhancing early and precise ALL diagnostics and reducing healthcare disparities in both high-income and low-and-middle-income countries.