Uncertainty Quantification in Segmenting Tuberculosis-Consistent Findings in Frontal Chest X-rays

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Deep learning (DL) excels in medical image segmentation but requires careful loss function selection for optimal performance.
• Traditional DL models lack uncertainty quantification, hindering trust in automated medical diagnostics.
• Quantifying prediction uncertainty is crucial for reliable medical applications.

Purpose of the Study:

• To investigate the benefits of an appropriate loss function and uncertainty quantification for segmenting tuberculosis (TB)-consistent findings in chest X-rays (CXRs).
• To evaluate a VGG16-based-U-Net model incorporating Monte Carlo (MC) Dropout for uncertainty estimation.
• To establish an optimal uncertainty threshold for referring uncertain cases to experts.

Main Methods:

• Utilized a VGG16-based-U-Net architecture for CXR segmentation.
• Employed a modified Focal Tversky loss function for improved segmentation performance.
• Implemented Monte Carlo Dropout with 30 forward passes for uncertainty quantification.
• Determined an optimal uncertainty threshold using various uncertainty metrics.

Main Results:

• The modified Focal Tversky loss function enhanced segmentation performance (mAP: 0.5710).
• Monte Carlo Dropout further improved and stabilized performance (mAP: 0.5721).
• An uncertainty threshold of 0.7 was identified as optimal for case referral.

Conclusions:

• The combination of an optimized loss function and uncertainty quantification significantly improves TB detection in CXRs.
• The proposed method provides a reliable mechanism for identifying and escalating uncertain cases to human experts.
• This approach enhances the trustworthiness and clinical utility of AI-driven diagnostic tools in radiology.