Multi-task interactive feature transfer network for dual-energy-like chest radiography image synthesis using CT data.

Accurate synthesis of bone and bone-suppression (B-S) images from chest radiographs is crucial for enhancing diagnostic precision in detecting lung abnormalities. Most deep learning methods, which typically generate only one type of image, limit diagnostic scope. This study introduces a novel multi-task interactive feature transfer network (MIFT-Net), which synthesizes dual-energy-like chest images simultaneously using computed tomography (CT) data. MIFT-Net employs a phased feature extraction and separation strategy to progressively generate clear bone and B-S images. By fusing features from different stages of the feature extraction process, the network promotes feature interaction and enhancement between the tasks of bone and B-S image synthesis. Finally, a multi-branch decoder outputs the generated images. For training, fast fuzzy C-means segmentation and pseudo x-ray techniques are utilized to generate bone and B-S images from CT data. MIFT-Net demonstrated superior performance on the test set, significantly outperforming existing image transformation methods in terms of peak signal-to-noise ratio, structural similarity index measure, and root-mean-square error for both bone and B-S images. A feasibility study on clinical chest x-ray images and a downstream diagnostic classification task suggest MIFT-Net's robustness and transferability, indicating its potential to enhance diagnostic accuracy and aid in clinical decision-making.