Don’t be wed to your classification model: deep learning-based models to diagnose idiopathic pulmonary fibrosis using computed tomography

<bold>Background:</bold> Machine learning and deep learning (DL) models have been developed to quantify or classify interstitial lung disease (ILD) patterns on high-resolution CT images (HRCT). A multi-scale guided attention (MSGA) DL was initially developed to distinguish idiopathic pulmonary fibrosis (IPF) from non-IPF ILD. <bold>Aims:</bold> This study aimed to explore using a MSGA DL model for this classification task and to investigate the implications of using the channel outputs and the attention mechanism from the DL model. <bold>Methods:</bold> A total of 878 volumetric non-contrast HRCT scans were retrospectively collected from IPF (N = 349, 39.7%) and non-IPF ILD cohorts (N = 529, 60.3%). A five-fold cross-validation was employed for training and validation of the MSGA DL model with a stratification of IPF and non-IPF subjects. Domain Knowledge is acquired as follows: (1) Voxel-level prediction: using the quantitative lung fibrosis (QLF) algorithm on CT scans, a prediction was made on a voxel-level; and (2) Population-level standardization. We provide both qualitative and quantitative methods to examine the extent of explainability. <bold>Results:</bold> Qualitatively, the observed attention maps with no or low attentions are uninformative and lack explainability. Quantitatively, the three best performing models had the AUC (±SD) values 0.987 (±0.007), 0.975 (±0.011), and 0.980 (±0.018), ranging from high to moderate MSGA with random forest classifier, respectively. <bold>Conclusions:</bold> A two-stage model for the IPF diagnosis among subjects with ILD based on chest HRCT was explored. The model combines an attention mechanism, for explainability and a random forest model for enhancing accuracy.