The Preliminary Study of Spectral CT and Machine Learning Method in Identifying Serosa Invasion of Gastric Cancer

Objective: To evaluate the value of spectral CT and machine learning method in identifying serosa invasion of gastric cancer. Method: Total of 24 cases of gastric cancer who underwent dual-phasic scans(arterial phase(AP) and portal phase(PP)) with GSI mode on high-definition computed tomography were retrospectively enrolled in our study, including 8 patients in p T2, 4 patients in p T3, and 12 patients in p T4. 12 patients(p T4 patients) were classified as serosa positive group, and 12 patients(p T2 and p T3 patients) were classified as serosa negative group. The clinical information(e.g. sex, age) of these two groups were compared by using independent sample t test or chi square test. In addition, GE AW4.4 workstation was used for image post-processing, and the dual phase spectrum information of these two groups was obtained. Support Vector Machine Recursive Feature Elimination(SVM-RFE) algorithm was used to analyze the spectrum information of these two groups. Results: Among the clinical information, only tumor long axis and short axis had statistically significant difference between two groups(all P < 0.05). The accuracies of SVM-RFE were 87.5%~94.4%. The output features of SVM-RFE were fat(calcium)(PP), uricacid(calcium)(PP), calcium(iodine)(AP), water(calcium)(PP), and iodine(water)(PP). Conclusion: Tumor size, fat(calcium)(PP), uricacid(calcium)(PP), calcium(iodine)(AP), water(calcium)(PP), and iodine(water)(PP) were helpful for the diagnosis of gastric cancer serosa invasion.