Letter to the Editor: Self-supervised learning model integrates plain and contrast-enhanced CT for preoperatively identifying gangrenous cholecystitis: a multicenter retrospective cohort study

Dear Editor, We read the article published by Guo et al[1] with great interest. The authors applied self-supervised learning (SSL) combined with multi-modal image fusion to preoperative gangrenous cholecystitis (GC) identification and conducted multi-center retrospective validation, achieving encouraging performance metrics. Here, we would like to share some additional insights that might be valuable for reference. This study is compliant with the Transparency In The reporting of Artificial INtelligence (TITAN) guideline[2]. The paper describes that “axial slices containing the entire gallbladder were manually selected by radiologists” and features were extracted from these 2D slices. However, gangrenous changes in the gallbladder are often diffuse and three-dimensionally distributed in time and space. Using a single slice to represent the entire organ may lead to information loss and label noise. Previous studies have emphasized that volume-based data better reflect the pathological characteristics of organs[3].We suggest adopting either full-volume input or a multi-slice aggregation strategy, and comparing the performance and stability of 2D versus 3D models to assess the bias caused by information loss. The study used postoperative pathology as the gold standard, with imaging acquisition restricted to within 48 hours before surgery. Nevertheless, the course of Gangrenous Cholecystitis progresses rapidly – its pathological state can evolve within hours to days preoperatively[4], and complications such as pathological progression, gallbladder wall necrosis, or perforation may occur during this period. If there is a time lag between CT imaging and the intraoperative or postoperative pathological state, the images may no longer reflect the final pathological condition. It would be advisable to clearly analyze the distribution of time intervals from CT scan to surgery or pathology in the report, and conduct subgroup sensitivity analysis accordingly to evaluate the impact of such delays on model performance. Overall, Guo et al’s work provides a valuable exploratory approach for the preoperative identification of gangrenous cholecystitis, particularly with its innovation in self-supervised learning and multi-modal fusion. We appreciate the authors’ efforts in this study and would like to thank both the authors and the editorial team for their contributions to the field. Through this letter, we aim to share some academic suggestions with the authors and journal readers, and we look forward to the research team further refining the methodology and promoting clinical trials to verify the real-world effectiveness of the model in the future. Ethical approval None. Consent None.