Enhanced Risk Stratification of Gastrointestinal Stromal Tumors Through Cross-Modality Synthesis from CT to [¹⁸F]-FDG PET Images

Abstract

Risk stratification algorithms for gastrointestinal stromal tumors (GISTs) are mainly based on computed tomography (CT) data. Though [18F]-fluorodeoxyglucose positron emission tomography ([18F]-FDG PET) imaging may improve their performance, challenges in image interpretation in the gastrointestinal tract still limit the widespread integration of PET into routine clinical protocols, causing a poor availability of PET data to develop and train stratification models. To solve this issue, we propose to enrich existing [18F]-FDG PET GIST datasets with pseudo-images generated with a novel conditional PET generative adversarial network (CPGAN), which employs a weighted fusion of CT images and tumor masks, embedding also clinical data. As for GIST assessment, we propose the transformer-based multimodal network for GIST risk stratification (TMGRS), which is trained on the enriched dataset and exploits the properties of transformers to process simultaneously PET and CT images. The training and validation of the models were conducted on a multicenter dataset comprising 208 patients. In comparison with the existing stratification methods, CPGAN-synthesized PET images show a peak signal-to-noise ratio increased on average by 18% and improve risk stratification, which achieves a remarkable accuracy of 0.937 when TMGRS network is used. Results underscore the potential of CPGAN network in providing more reliable GIST predictions.