Evaluation of gastric cancer using artificial intelligence iterative reconstruction on abdominal CT: image quality and diagnostic accuracy.

PURPOSE: To investigate the clinical performance of a novel deep-learning based image reconstruction algorithm, namely artificial intelligence iterative reconstruction (AIIR), for assessing gastric cancer (GC) on CT. MATERIALS AND METHODS: We prospectively enrolled 132 GC patients without history of treatment, all of whom subsequently underwent surgical resection or staging laparoscopy. All patients underwent preoperative abdominal contrast-enhanced CT examinations, and images were reconstructed with both hybrid iterative reconstruction (HIR) and AIIR. Qualitative metrics, including conspicuity of tumor margin and enhancement pattern, were evaluated using a 5-point Likert scale (1: poor; 5: excellent). Quantitative image quality was evaluated by calculating the contrast-to-noise ratio (CNR) of GC. The diagnostic performance in detecting gastric serosal invasion was characterized using receiver operating characteristic (ROC) analysis with the operative reference standard. RESULTS: The mean effective dose for CT examination was 15.3 ± 4.5 mSv. Compared to HIR, AIIR showed superior conspicuity of tumor margin (4.51 ± 0.81 vs. 3.90 ± 0.89, p < 0.001) and enhancement pattern (4.48 ± 0.76 vs. 3.86 ± 0.84, p < 0.001). The CNR of GC was significantly higher on AIIR than HIR in both arterial and portal venous phases (both p < 0.001). Accordingly, AIIR achieved a significantly higher area under the ROC curve than HIR in detecting gastric serosal invasion [0.89 (95%CI: 0.83-0.94) vs. 0.78 (95%CI: 0.70-0.85), p< 0.001]. CONCLUSION: Compared to HIR, AIIR yields better image quality and was associated with better diagnostic performance for the assessment of gastric serosal invasion on routine abdominal CT, suggesting its potential clinical value in the preoperative evaluation of GC.