A Morphological Context Blocks Hybrid CNN for Efficient Acute Lymphoblastic Leukemia Classification

Acute Lymphoblastic Leukemia (ALL) is an aggressive hematologic malignancy that necessitates early and accurate diagnosis for improved therapeutic efficacy. Although it is a routine practice, the visual blood smear analysis is tedious and subject to human inaccuracies. This paper proposes a novel morphology-guided deep learning approach called Morphological Context Blocks (MCB)-HyperNet embedding morphological operations into a hybrid CNN architecture. The CNN architectures depend mainly on automatic learning through convolutive filters, so they miss crucial morphological features that distinguish between leukemic and normal cells. In this study, we propose a deep learning-based approach that directly incorporates morphological dilation and erosion in the deep learning data pipeline to exploit the potential of morphological feature extraction for our specific task, resulting in enhanced accuracy and reduced diagnostic costs, which ultimately can improve patient outcomes. In addition, the computational efficiency and modularity of the MCB-HyperNet framework facilitate easy adaptation and scalability to many other medical imaging tasks, such as the classification of various diseases, except the classification of leukemia. We trained the proposed MCB-HyperNet on different image resolutions from the ALL dataset (168×168, 224×224, 256×256), different batch sizes (16 and 32), and also different training epochs (30, 35, 40, 45, 50) to get the best hyperparameter configuration. The MCB-HyperNet takes advantage of the strong feature extraction ability of ResNet and the light computing resource of MobileNetV3, ultimately obtaining 99.69% accuracy, 98.78% precision, 99.49% sensitivity, 99.12% F1-score, and 99.78% specificity. This new integration greatly enhances the accuracy of early detection, minimizes diagnostic errors, and could have significant clinical and economic advantages. MCB-HyperNet is a mini CNN, so it shows a good balance between efficiency and accuracy, making scalability and extensibility possible in more medical imaging tasks.