An uncertainty-aware vision transformer-BiLSTM Bayesian framework for reliable clinical decision support using chest...

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Deep Learning for Diagnostics

Background:

• Deep learning models in medical imaging face challenges in reliability and uncertainty estimation.
• Conventional Convolutional Neural Network (CNN)-based models often exhibit extreme prediction confidence, limiting clinical trust.
• Accurate uncertainty quantification is crucial for high-stakes diagnostic decisions in clinical practice.

Purpose of the Study:

• To develop an uncertainty-aware deep learning framework for reliable disease detection using chest X-rays.
• To enhance the interpretability and trustworthiness of AI models in clinical decision support.
• To address limitations in current deep learning models regarding reliability and confidence estimation.

Main Methods:

• Proposed a Vision Transformer-BiLSTM Bayesian Fusion (ViT-BiLSTM-BF) architecture.
• Utilized chest X-rays from MIMIC-CXR-JPG and PadChest-GR datasets for training and evaluation.
• Incorporated U-Net based lung segmentation, data augmentation, and a Bayesian fusion layer for uncertainty quantification.

Main Results:

• Achieved superior performance with 95.5% accuracy and 0.986 AUC-ROC.
• Demonstrated extremely small calibration error (ECE = 0.028), indicating well-calibrated confidence measures.
• Outperformed existing models including CNN, DenseNet, Bayesian CNN, and individual ViT models.

Conclusions:

• The ViT-BiLSTM-BF model offers a highly reliable and clinically meaningful diagnostic system for chest X-ray analysis.
• The framework effectively addresses critical issues of uncertainty estimation and model reliability in medical AI.
• The proposed approach shows potential for feasible implementation in clinical decision support systems.