Leveraging Transfer Learning and Monte Carlo Dropout for Uncertainty Informed NIRS-based Detection of Systemic...

Area of Science:

• Biomedical optics
• Medical imaging analysis
• Artificial intelligence in healthcare

Background:

• Early diagnosis of systemic sclerosis (SSc) is crucial for timely intervention and improved patient prognosis.
• Microvascular dysfunction is a hallmark of SSc, necessitating reliable detection methods.
• Current diagnostic approaches may lack noninvasiveness or objectivity for microvascular assessment.

Purpose of the Study:

• To investigate the efficacy of integrating near-infrared spectroscopy (NIRS) with deep learning for classifying SSc patients.
• To analyze hand perfusion patterns using NIRS-derived oxygen saturation maps for SSc detection.
• To assess the reliability and robustness of the deep learning model through uncertainty estimation.

Main Methods:

• Utilized near-infrared spectroscopy (NIRS) to acquire oxygen saturation maps of the hand.
• Employed a probabilistic convolutional neural network (CNN) with MobileNetV2 and transfer learning for classification.
• Incorporated Monte Carlo Dropout for evaluating predictive uncertainty and detecting out-of-distribution inputs.

Main Results:

• The deep learning model achieved a test accuracy of 87.5% in classifying SSc patients.
• Analysis of confidence levels across datasets highlighted the importance of uncertainty estimation for model reliability.
• Demonstrated feasibility of NIRS combined with deep learning for detecting microvascular dysfunction in SSc.

Conclusions:

• Deep learning-based NIRS analysis presents a viable noninvasive, automated tool for SSc detection.
• The study confirms the potential for objective assessment of microvascular changes in SSc.
• Future research should focus on larger datasets and multimodal integration for enhanced clinical applicability.