Tuberculosis detection using few shot learning

Tuberculosis (TB), a contagious disease, significantly affects lungs functioning. Amongst multiple detection methodologies, Chest X-ray analysis is considered the most effective methodology. Traditional Deep Learning methodologies have shown good results for TB detection; however, model's huge number of parameters, size, and compute requirements making it unsuitable for its practical deployment. Owing to scarce annotated datasets in medical domain augmented datasets are generated which is not a recommended technique in medical domain. This study presents TB-FSNet consisting of Few Shot Learning - Prototypical Network (FSL-PT) with a modified MobileNet-V2 backbone, incorporating a Self-Attention layer. The significant contribution of this study is to effectively train TB-FSNet in FSL-PT paradigm with six different backbones. The dataset utilised for this study consists of Montgomery County, and Shenzhen Chest X-ray Dataset combined. The proposed method attains highest accuracy of 93.6% with mere 2.21M parameters and 8.67 MB size, while maintaining high performance metrics such as precision, specificity, and sensitivity. Moreover, TB-FSNet is designed for seamless integration into embedded devices, making it suitable for deployment on edge devises. The model processes Chest X-ray images in real-time, providing immediate confidence scores for disease detection. This capability ensures that users can receive accurate diagnostic insights without needing to wait for medical professionals, enhancing the accessibility and efficiency of TB detection.