A novel dual-branch CNN architecture for explainable feature agnostic predictions in asthma

Résumé fourni par la source

Background – Use of computer vision in chronic airway disease management remains limited and often relies solely on inspiratory scans. Objective – Design and test a novel convolutional neural network (CNN) architecture for application to multiple asthma labels. Method – Inspiratory and expiratory scans from the (Assessment of Small Airways Involvement in Asthma) ATLANTIS ( NCT02123667 ) cohort (310 asthma, 51 healthy subjects) were resized via bicubic or spline interpolation for upsampling and Gaussian pyramid for downsampling, with Structural Similarity Index (SSIM) assessing efficacy. Images were clipped between -400 and 1024 HU, followed by three-channel preprocessing (binarisation, edge detection, dilation). Two volumetric CNNs processed inspiratory/expiratory scans separately, featuring interchangeable backbones (e.g., ResNet18, V-Net). Outputs were concatenated and fed into a classifier. Full-scan reconstruction and prediction area localisation were conducted using Grad-CAM++ mapping. Results – Resizing was effective, with an SSIM of 0.917 for spline interpolation upsampling and 0.863 for Gaussian pyramid downsampling. Images with three-channel preprocessing significantly outperformed those with simple clipping and normalisation in a 2D Long Short-Term Memory CNN for predicting GINA severity. Slice reconstruction and Grad-CAM++ mapping was successfully applied, enhancing interpretability. Conclusion – Preliminary testing of components of the CNN architecture has yielded promising results. Once fully optimised, it could be used to map CT scans to various asthma labels including GINA and lung function, with potential extension to multiple labels and multi-modal data integration.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.