Attention-boosted deep learning model distinguishes IPF from other interstitial diseases on CT

In a retrospective study of 96 patients with UIP pattern on HRCT, an attention-enhanced VGG-16 model distinguished IPF from non-IPF with 87.2% accuracy, AUC 0.91, sensitivity 86.1%, specificity 88.4%. The model highlighted subpleural reticular opacities and honeycombing.

• Retrospective, single-center study of 96 patients (46 IPF, 50 non-IPF) with UIP pattern on HRCT.
• SE-VGG-16 model achieved accuracy 87.2%, AUC 0.91, and outperformed baseline VGG-16 by 4.1%.
• Grad-CAM heatmaps consistently highlighted subpleural reticular opacities and honeycombing.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.