Novel multiparametric MRI-based radiomics in preoperative prediction of perirectal fat invasion in rectal cancer

• Published in: Abdominal imaging Vol. 48; no. 2; pp. 471 - 485
• Main Authors: Wang, Hui, Chen, Xiaoyong, Ding, Jingfeng, Deng, Shuitang, Mao, Guoqun, Tian, Shuyuan, Zhu, Xiandi, Ao, Weiqun
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.02.2023; Springer Nature B.V
• Subjects: Accuracy; Biomarkers; Calibration; Cancer; Colorectal cancer; Decision analysis; Gastroenterology; Hepatology; Hollow Organ Gi; Humans; Image enhancement; Imaging; Magnetic Resonance Imaging; Medical imaging; Medicine; Medicine & Public Health; Model accuracy; Multiparametric Magnetic Resonance Imaging; Nomograms; Radiology; Radiomics; Rectal Neoplasms - diagnostic imaging; Rectal Neoplasms - surgery; Rectum; Retrospective Studies; Statistical analysis; Statistics; Training; Perirectal fat invasion; Rectal cancer; Magnetic resonance imaging; Nomogram; Radiomics
• ISSN: 2366-0058; 2366-004X; 2366-0058
• DOI: 10.1007/s00261-022-03759-z

Objectives To investigate the feasibility and efficacy of a nomogram that combines clinical and radiomic features of magnetic resonance imaging (MRI) for preoperative perirectal fat invasion (PFI) prediction in rectal cancer. Methods This was a retrospective study. A total of 363 patients from two centers were included in the study. Patients in the first center were randomly divided into training cohort ( n  = 212) and internal validation cohort ( n  = 91) at the ratio of 7:3. Patients in the second center were allocated to the external validation cohort ( n  = 60). Among the training cohort, the numbers of patients who were PFI positive and PFI negative were 108 and 104, respectively. The radiomics features of preoperative T 2 -weighted images, diffusion-weighted images and enhanced T 1 -weighted images were extracted, and the total Radscore of each patient was obtained. We created Clinic model and Radscore model, respectively, according to clinical data or Radscore only. And that, we assembled the combined model using the clinical data and Radscore. We used DeLong’s test, receiver operating characteristic, calibration and decision curve analysis to assess the models’ performance. Results The three models had good performance. Clinic model and Radscore model showed equivalent performance with AUCs of 0.85, 0.82 (accuracy of 81%, 81%) in the training cohort, AUCs of 0.78, 0.86 (accuracy of 74%, 84%) in the internal cohort, and 0.84, 0.84 (accuracy of 80%, 82%) in the external cohort without statistical difference (DeLong’s test, p  > 0.05). AUCs and accuracy of Combined model were 0.89 and 87%, 0.90 and 88%, and 0.90 and 88% in the three cohorts, respectively, which were higher than that of Clinic model and Radscore model, but only in the training cohort with a statistical difference (DeLong’s test, p  < 0.05). The calibration curves of the nomogram exhibited acceptable consistency, and the decision curve analysis indicated higher net benefit in clinical practice. Conclusion A nomogram combining clinical and radiomic features of MRI to compute the probability of PFI in rectal cancer was developed and validated. It has the potential to serve as a preoperative biomarker for predicting pathological PFI of rectal cancer. Graphical abstract