A Comparative Study of Acute Lymphoblastic Leukemia Diagnosis Using CNN, Swin-Transformer, and SSM-Based AI Approaches

The traditional diagnostic techniques for Acute Lymphoblastic Leukemia (ALL) are constrained by technical limitations and subjective assessments, leading to a misdiagnosis rate of 15 % to 25 %. Thus, early and precise diagnosis is crucial for advancing precision medicine and minimizing the costs associated with early treatment. This study seeks to develop an efficient intelligent diagnostic model to address the high rate of missed diagnoses and the limitations in feature extraction inherent in conventional medical imaging diagnostics. The objective is to achieve accurate classification of ALL and its subtypes. Utilizing an open-source dataset from Kaggle, comprising 3,256 Peripheral Blood Smear (PBS) images, the dataset was partitioned into training, validation, and test sets in a 6: 2: 2 ratio. To enhance data randomness and diversity, preprocessing techniques such as contrast-limited adaptive histogram equalization and median filtering were employed. The study evaluates the performance of seven Convolutional Neural Network (CNN) models and the Swin Transformer model, with a particular focus on the novel integration of the MedMamba model for its potential advantages in processing medical images. The investigation into deep learning models, particularly those utilizing pre-trained weights like the Swin Transformer model, demonstrates a substantial improvement in the recognition accuracy of ALL (achieving a rate of 99.23 %), thereby mitigating the additional medical costs and ethical concerns associated with misdiagnoses or missed diagnoses.