Abstract

This study systematically evaluates four MobileNet architectures MobileNetV1, MobileNetV2, MobileNetV3, and MobileNetV4 for the segmentation of tuberculosis (TB) lesions in chest X-ray images. We focus on the critical balance between segmentation accuracy and computational efficiency, assessing model performance using key metrics including F1-Score, mean Intersection over Union (mIoU) and recall. Computational needs are measured in terms of the number of parameters, floating-point operations (FLOPs), and inference time. We have observed that there exists a significant trade-off: MobileNetV2 is the fastest in terms of segmentation accuracy and can be chosen when precision is a necessity. In the meantime, MobileNetV3 is more successful with lowest possible parameter count, and is therefore especially well adapted to resources constrained environments (like mobile devices, real time clinical settings). These findings underline the fact that the choice of models must be guided by the individual requirements and limitations of the target application. Future studies will aim at testing these architectures with more extensive and more heterogeneous clinical data, and explore superior ways to optimize the models in order to find a better balance between accuracy and efficiency, and eventually improve the creation of affordable and reliable tuberculosis diagnostic models.