Vision Mamba for efficient Tuberculosis Detection based on Chest X-Rays: A comparative study with CNN and Vision transformers

The rapid growth of medical images creates substantial challenge for manual interpretation and classification. Computer-aided diagnostic systems address this challenge leveraging different deep learning models to assist in image analysis. Convolutional Neural Networks process images with local context and struggle in capturing long-range context. Vision transformers address this limitation effectively. Whereas, the quadratic computational complexity caused by the attention mechanism has proven to be an obstacle to further scaling such models on edge and low-resource devices. Recently, Vision Mamba-based-models, an alternative architecture to transformers, have emerged. They excel in capturing local and long-range dependencies through the bidirectional Mamba block, while maintaining linear computational complexity. In this study, we fine-tuned a Vision Mamba model for Tuberculosis detection using chest X-ray images extracted from a merged dataset. The proposed Vision Mamba model achieves a valuable accuracy of 94.32% outperforming traditional CNNs maintaining competitive performance with ViT models. It reduces GPU memory usage by 80%, addressing key challenges in deploying deep learning models in resource-constrained environments. Vision Mamba demonstrates strong potential for clinical application, particularly in automating tuberculosis diagnosis in regions with limited computational resources or medical expertise.