Predicting Usual Interstitial Pneumonia Histopathology From Chest CT Imaging With Deep Learning.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Pulmonary Medicine
• Radiology

Background:

• Idiopathic pulmonary fibrosis (IPF) is a progressive, fatal interstitial lung disease (ILD) requiring accurate diagnosis.
• Distinguishing usual interstitial pneumonitis (UIP)/IPF from other ILDs is crucial for treatment and prognosis.
• Lung biopsy is often necessary when noninvasive data are insufficient for diagnosis.

Purpose of the Study:

• To investigate if deep learning can improve the noninvasive diagnosis of UIP/IPF by predicting ILD histopathology from CT scans.
• To compare the diagnostic performance of a deep learning model against expert radiologists.

Main Methods:

• Retrospective analysis of 1,239 patients with pathologically proven ILD and chest CT scans.
• A deep learning model was trained to predict UIP/non-UIP histopathology from CT images.
• Model performance was compared to manual CT scan labeling by two radiologists using ROC analysis.

Main Results:

• Deep learning model achieved superior performance in predicting histopathologic diagnosis (AUC 0.87) compared to visual analysis by radiologists (AUC 0.80, P < .05).
• The deep learning model demonstrated significantly greater reproducibility than both inter-rater and intra-rater radiologist reproducibility.
• The study included a cohort of 1,239 patients, with the model evaluated on a 198-patient test set.

Conclusions:

• Deep learning models show potential to outperform visual assessment in predicting UIP/IPF histopathology from CT imaging.
• This AI-driven approach may offer a noninvasive alternative to lung biopsy for diagnosing UIP/IPF.
• Further validation is warranted to integrate deep learning into clinical practice for ILD diagnosis.