Automated diagnosis of usual interstitial pneumonia on chest CT via the mean curvature of isophotes

Purpose To test whether the mean curvature of isophotes (MCI), a geometric image transformation, can be used to improve automatic detection on chest CT of Usual Interstitial Pneumonia (UIP), a determining radiological pattern in the diagnosis of Interstitial Lung Diseases (ILD).

Materials and methods This retrospective study included chest CT scans from 234 patients (123 female,111 male; mean age: 61.6 years; age range: 18-90 years) obtained at two independent institutions between 2007 and 2024.

Three different classification models were trained on the original CT images and separately on MCI-transformed CT images: (1) a previously published deep learning model for classifying fibrotic lung disease on chest CT, (2) a simplified version of the same, and (3) a non-deep-learning model based on the functional principal component analysis (FPCA) of density functions of voxel intensity.

All models were trained on data from the first institution and evaluated on data from the second institution with the recall-macro, precision-macro and F1-macro scores. Performance difference between classifier pairs was tested with the Stuart-Maxwell marginal homogeneity test.

Results For a fixed model architecture and training algorithm, MCI-transformed images yield comparable or better classification performance than the original CT images. The best performance improvement achieved with MCI compared to CT was: recall-macro 0.83 vs 0.57, precision-macro 0.81 vs 0.50, F1-macro 0.80 vs 0.49, p=4.2e-5.

Conclusion MCI may be a valuable addition to existing AI systems for screening for UIP on chest CT.

Summary statement Machine learning methods for identifying usual interstitial pneumonia on chest CT perform better when the input CT images are transformed via the mean curvature of isophotes (MCI), a geometric transformation method known from classical computer vision.

Key Points

• Three machine learning models were trained on a dataset of 158 patients from one institution and tested on another dataset of 76 patients from an independent institution to discriminate for usual interstitial pneumonia (UIP) on chest CT in a 3-group classification task.

• When keeping the network architecture and parameters fixed, changing the input image domain from the original CT to MCI-transformed images improved classification performance (Stuart-Maxwell test, p < .001)

• MCI may be a valuable addition to existing machine learning systems for screening for UIP on chest CT, whether based on deep learning or on simpler shallow classifiers.