Scholarlink

Background <p> Each year, pneumonia is responsible for more than 2.2 million deaths. Although deep learning has helped to analyze chest X-rays (CXRs), current systems approach segmentation and classification separately. They also use basic combination methods and give clinically inappropriate overconfident predictions.</p> Purpose <p>To formulate and verify a pneumonia detection framework with the clinically safe deployment of anatomically informed segmentation, cross-attention fusion, and ensemble calibration.</p> Methods <p>We created a MedSAM-guided Transformer enhanced U-Net segmentation framework combined with cross attention from dual CheXFound branch encoders. Training was done on the RSNA and NIH ChestX-ray14 datasets (n=28,526), and external validation was done on the CheXpert and COVIDx datasets. The ensembles of five models with temperature scaling improved calibration. Key metrics to evaluate the performance of the models include AUROC, expected calibration error (ECE), and radiologist Grad-CAM++ annotations for interpretability validation.</p> Results <p>On combined test sets (n=6,112), the framework achieved AUROC 0.973 [95% CI: 0.969–0.977], significantly outperforming ResNet-50 (0.921), DenseNet-121 (0.934), ViT-B/16 (0.947), and CheXFound (0.958; all p&lt;0.05). External validation demonstrated robust generalization (CheXpert: 0.961; COVIDx: 0.958). Cross-attention fusion outperformed concatenation (0.973 vs. 0.963, p=0.008). Ensemble calibration reduced ECE by 62% (0.021→0.008). Grad-CAM++ showed 87.3% overlap with radiologist annotations.</p> Conclusions <p>This calibration-aware framework demonstrates clinically relevant performance through unified integration of segmentation, validated interpretability, and trustworthy probability estimation, advancing development of safe AI-assisted pneumonia screening systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Calibrated pneumonia detection in chest X-rays via cross-attention fusion of foundation model features and lung-masked representations

  • 期刊论文

  • 发表日期：

    2026年4月17日

查看全文
  • Honey Joseph,
  • Anitha J

摘要

Background

Each year, pneumonia is responsible for more than 2.2 million deaths. Although deep learning has helped to analyze chest X-rays (CXRs), current systems approach segmentation and classification separately. They also use basic combination methods and give clinically inappropriate overconfident predictions.

Purpose

To formulate and verify a pneumonia detection framework with the clinically safe deployment of anatomically informed segmentation, cross-attention fusion, and ensemble calibration.

Methods

We created a MedSAM-guided Transformer enhanced U-Net segmentation framework combined with cross attention from dual CheXFound branch encoders. Training was done on the RSNA and NIH ChestX-ray14 datasets (n=28,526), and external validation was done on the CheXpert and COVIDx datasets. The ensembles of five models with temperature scaling improved calibration. Key metrics to evaluate the performance of the models include AUROC, expected calibration error (ECE), and radiologist Grad-CAM++ annotations for interpretability validation.

Results

On combined test sets (n=6,112), the framework achieved AUROC 0.973 [95% CI: 0.969–0.977], significantly outperforming ResNet-50 (0.921), DenseNet-121 (0.934), ViT-B/16 (0.947), and CheXFound (0.958; all p<0.05). External validation demonstrated robust generalization (CheXpert: 0.961; COVIDx: 0.958). Cross-attention fusion outperformed concatenation (0.973 vs. 0.963, p=0.008). Ensemble calibration reduced ECE by 62% (0.021→0.008). Grad-CAM++ showed 87.3% overlap with radiologist annotations.

Conclusions

This calibration-aware framework demonstrates clinically relevant performance through unified integration of segmentation, validated interpretability, and trustworthy probability estimation, advancing development of safe AI-assisted pneumonia screening systems.