TOPLINE:

Convolutional neural networks (CNNs) trained on lung ultrasound (LUS) images detected interstitial lung disease (ILD) in patients with systemic sclerosis and idiopathic inflammatory myopathy, showing agreement with human interpretation.

METHODOLOGY:

• Researchers conducted a retrospective study to evaluate whether CNNs could accurately detect ILD and assess its severity using LUS images.
• They analyzed 3920 LUS images from 140 adult patients with systemic sclerosis or idiopathic inflammatory myopathy who had undergone both LUS and chest CT scans, dividing the dataset into a development set (n = 74) and an independent test set (n = 66; mean age, 56 years; 79% female).
• Three pretrained CNN architectures (InceptionV3, Residual Network-50, and Visual Geometry Group [VGG]-16) were trained using transfer learning, and a novel lightweight model (LUS-Net) was developed to detect ILD-positive and ILD-negative LUS images.
• Model performance was evaluated at both image and patient levels in the independent test set using receiver operating characteristic curves, sensitivity, specificity, and agreement with expert interpretation.
• Correlations between CNN outputs (from VGG-16 and LUS-Net), LUS-ILD-24 scores (for human interpretation), pulmonary function test results, and CT-based CALIPER indices were assessed. Gradient-weighted class activation mapping (Grad-CAM) was used to visualize the LUS regions that contributed most to CNN predictions.

TAKEAWAY:

• The VGG-16 model achieved the best performance for ILD detection at the patient level, with an area under the curve (AUC) of 0.972, sensitivity of 97.4%, and specificity of 92.6%, and showed agreement with expert human interpretation (Cohen kappa, 0.804; 95% CI, 0.654-0.953).
• The CNN outputs showed strong correlations with human interpretation and CALIPER total percent ILD, moderate-to-strong correlations with CALIPER total percent fibrosis, and moderate negative correlations with diffusion capacity of the lungs for carbon monoxide (for VGG-16, Spearman rank correlation coefficients, 0.931, 0.728, 0.595, and -0.526, respectively).
• At the image level, the VGG-16 model detected ILD with an AUC of 0.866 and showed agreement with consensus human interpretation (Cohen kappa, 0.62; 95% CI, 0.56-0.67).
• Grad-CAM analysis showed that both VGG-16 and LUS-Net focused primarily on pleural surface features and the subpleural echogenic gradient, identifying activations in ILD-positive images even in the absence of B-lines by highlighting pleural irregularity, granularity, and pseudo-thickening.

IN PRACTICE:

“By combining the portability and safety of LUS with the power of CNNs, this [deep learning-based] approach has the potential to standardize interpretation, enhance understanding and transparency with explainable AI, and expand access to high-quality ILD screening and monitoring across a broad range of clinical settings,” the authors of the study wrote.

SOURCE:

The study was led by Robert M. Fairchild, MD, PhD, Stanford University School of Medicine, Palo Alto, California. It was published online on April 6, 2026, in Arthritis Care & Research.

LIMITATIONS:

The study used LUS data from a single platform and acquired by experienced operators, which may not reflect broader clinical variability. The sample size was modest compared with large-scale deep learning benchmarks. The model performance was lower for individual image classification than for aggregate patient-level assessment.

DISCLOSURES:

Two authors reported receiving funding from Boehringer Ingelheim, and one of them also received support from the Scleroderma Research Foundation. One author reported receiving consulting fees, honoraria, or support for meetings and/or travel from various pharmaceutical companies or organizations; another author reported receiving a grant; and a third author disclosed receiving grants or contracts and consulting fees, participation on a data safety monitoring board or advisory board, and stock ownership in various companies.

This article was created using several editorial tools, including AI, as part of the process. Human editors reviewed this content before publication.