Method for explaining deep learning model predictions of lung cancer risk based on computed tomography images using image-to-image Schr dinger Bridge generative reconstruction

Abstract

The subject of the application is an advanced medical image analysis system, shown in the figure, that enables the generation and archiving of diagnostic results in DICOM format. The system is based on a modular architecture, in which a local reconstruction module plays a special role, utilizing a U-Net generative model with attention mechanisms. By using the Schrödinger bridge trajectory technique, precise image modifications can be performed in designated regions of interest (ROIs). The system outputs not only original and counterfactually modified images but also ROI masks and input values, which can be archived in a PACS system.The system enables automatic processing of image data and ensures high-quality reconstruction thanks to advanced machine learning algorithms. The solution is applicable in radiology, oncology, and other fields of imaging medicine where precise identification and analysis of anatomical structures are essential. The application also concerns a method for computer-aided explanation of lung cancer risk model predictions based on volumetric computed tomography images.