BacilliFinder: Revolutionizing Tuberculosis Detection with Computer Vision

Background Pulmonary tuberculosis, caused by Mycobacterium, remains a global health concern despite being treatable, preventable, and curable. Diagnosis traditionally relies on X-rays and sputum smear samples. However, sputum smear tests are dependent on the quality of sputum, and X-rays often lack the necessary contrast. Both methods demand skilled technicians and require a closer examination. Methods This paper introduces an innovative solution, presenting an efficient and computationally lighter deep convolutional model for the detection of Mycobacterium Bacilli in sputum smear samples. The model is deliberately downscaled in width and depth to reduce computational demands, enabling its integration into edge devices. Furthermore, to make the model more robust, a collection of sputum smear test sample images dataset is formed for this study. Results The performance of the proposed model is assessed using our own dataset. The results demonstrate exceptional stability and a significantly improved detection rate. Specifically, the model achieves an impressive mean Average Precision (mAP) of 87.2 % at an Intersection over Union (IoU) threshold of 0.5 and a mAP of 50.5 % at an IoU threshold ranging from 0.5 to 0.95. Additionally, the model exhibits robust recall and precision metrics, standing at 86 % and 84 %, respectively. Conclusions The detection results indicate that the proposed YOLOv7_Lite model addresses challenges associated with close examination and mitigates potential biases. Its efficiency and adaptability to memory-constrained devices make it a valuable tool that can significantly contribute to early and accurate tuberculosis diagnosis, by increasing the sensitivity of smear microscopy.