A Deep Learning-Based Radiomic Classifier for Usual Interstitial Pneumonia.

Area of Science:

• Radiology
• Artificial Intelligence
• Pulmonology

Background:

• Chest CT scans are primary tools for diagnosing interstitial lung disease (ILD), replacing surgical lung biopsy.
• Standardized interpretation of CT scans is crucial for accurate ILD diagnosis.
• Deep learning (DL) offers potential for automating and standardizing CT scan analysis.

Purpose of the Study:

• To develop and evaluate a DL-based classifier for usual interstitial pneumonia (UIP) using chest CT features.
• To assess the accuracy of the DL classifier in discriminating radiologist-determined visual UIP.
• To compare the survival discrimination of DL-based and visual UIP classification.

Main Methods:

• Retrospective study of 2,907 chest CT scans from diverse data sources.
• A convolutional neural network trained on CT features to predict UIP likelihood.
• Linear support vector machine used for classification, with performance assessed in independent cohorts.
• Transplant-free survival analyzed using Kaplan-Meier and Cox regression.

Main Results:

• The DL-based UIP classifier achieved 93% sensitivity and 86% specificity in the performance cohort.
• In a multicenter ILD cohort, the classifier showed 81% sensitivity and 77% specificity.
• DL-based and visual UIP classifications demonstrated similar survival discrimination.
• Outcomes were consistent for positive DL-based UIP classification regardless of visual classification.

Conclusions:

• A DL-based classifier for UIP demonstrates robust performance across varying prevalence.
• The automated tool accurately identifies UIP on chest CT scans, comparable to expert radiologists.
• This DL classifier can efficiently screen for UIP and identify high-risk ILD phenotypes.