Abstract
Survival prediction models for COVID-19 are crucial for clinical triage but often lack interpretability. We present a deep learning framework that combines medical vision-language pretraining with a novel "Region Completer" to provide region-specific risk quantification from chest X-rays. By anchoring risk-specific Grad-CAM heatmaps to 29 anatomical regions, our model achieves high prognostic accuracy (C-index 0.764) while offering clinicians a transparent, ranked list of high-risk anatomical zones.
Introduction
In the heat of the COVID-19 pandemic, survival prediction models became vital for triaging patients and managing depleted healthcare resources. However, most deep learning models suffer from the "black box" problem: they might accurately predict a patient's risk, but they cannot explain why or where they see that risk in a medical image. For a radiologist, a single risk score is often insufficient; they need to know if the model is focusing on relevant pulmonary manifestations or irrelevant artifacts. This paper aims to bridge the gap between high-performance AI and clinical trust by providing region-specific risk quantification.
Methodology
The proposed pipeline is designed for both accuracy and transparency through several key stages:
- Medical Vision-Language Pretraining (PRIOR): The model utilizes the PRIOR encoder, pretrained on the MIMIC-CXR dataset. This provides a domain-specific foundation for understanding chest anatomy compared to standard ImageNet pretraining.
- Survival Prediction: A ResNet50 backbone coupled with a survival predictor is trained using the Cox Proportional Hazards (CoxPH) loss, effectively handling censored survival data.
- Anatomical Region Detection and Completion: To ensure robust localization, the system uses a Faster R-CNN detector followed by a novel "Region Completer." This regression network uses spatial patterns of detected boxes to infer the locations of missing regions, ensuring a full 29-region anatomical workup for every image.
- Regional Risk Scoring: Risk-specific Grad-CAM heatmaps are generated by backpropagating the risk score. These heatmaps are then intersected with the 29 anatomical boxes to calculate a localized risk score for each region.
Key Results
The model was evaluated using the Concordance Index (C-index) and Time-Dependent AUC (T-AUC) on multi-center datasets.
- Performance: The model achieved a C-index of 0.764 on the internal test set and 0.727 on an external multi-center set, outperforming traditional methods like Random Survival Forests and standard CoxPH models.
- Interpretability: Qualitative analysis shows the model successfully localizes ground-glass opacities and ranks affected lung zones as high-risk areas, aligning with clinical reasoning.
Discussion and Conclusion
The core innovation of this work is anchoring interpretability to anatomy. A heatmap alone is often ambiguous, but a "Risk Score of 0.745 in the Left Lower Lung" provides a concrete clinical finding. The "Region Completer" further enhances reliability by handling the inherent noise and occlusion in diseased chest X-rays.
While currently validated for COVID-19, the framework is generalizable to other pulmonary diseases. Future work should focus on clinical integration studies to assess how this regional data impacts real-world diagnostic speed and accuracy.
Summary updated