Semantic Segmentation of TB in Chest X-rays: a New Dataset and Generalization Evaluation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Public Health

Background:

• Tuberculosis (TB) diagnosis relies heavily on chest X-rays (CXRs), but automated methods lack explainability.
• Semantic segmentation of TB lesions in CXRs offers a path toward enhanced human oversight in diagnosis.
• The global burden of TB necessitates improved and transparent diagnostic tools.

Purpose of the Study:

• To introduce the TB-Portals SIFT dataset for semantic segmentation of TB lesions in CXRs.
• To evaluate deep learning models for TB lesion segmentation and compare their generalization capabilities.
• To assess the performance of segmentation-based classifiers against traditional image-level classifiers.

Main Methods:

• Developed the TB-Portals SIFT dataset with 6,328 CXRs and 10,435 pseudo-labeled TB lesion instances.
• Evaluated ten semantic segmentation models (UNet and YOLOv8-seg architectures) using five-fold cross-validation.
• Compared the generalization of top segmentation models and their ensemble against DenseNet121 classifiers on classification and object detection tasks.

Main Results:

• The segmentation-based approach demonstrated superior generalizability compared to DenseNet121 classifiers.
• An ensemble of UNet and YOLOv8-seg models achieved the most stable performance, excelling across segmentation, classification, and object detection.
• The best segmentation models from each architecture (nnUNet(ResEnc XL) and YOLOv8m-seg) were identified.

Conclusions:

• Semantic segmentation of TB lesions in CXRs is a promising approach for improving diagnostic accuracy and explainability.
• The TB-Portals SIFT dataset and evaluated models offer valuable resources for advancing automated TB detection.
• Ensemble methods combining different architectures provide robust and generalizable solutions for TB analysis in medical imaging.