Abstracts

RETURN TO ABSTRACT LISTING


2096. MRI Radiomics Features of Mesorectal Fat Can Predict Prognosis in Patients with Locally Advanced Rectal Cancer
Authors * Denotes Presenting Author
  1. Vetri Sudar Jayaprakasam *; Memorial Sloan Kettering Cancer Centre
  2. Viktoriya Paroder; Memorial Sloan Kettering Cancer Centre
  3. Peter Gibbs; Memorial Sloan Kettering Cancer Centre
  4. Raazi Bajwa; Memorial Sloan Kettering Cancer Centre
  5. Natalie Gangai; Memorial Sloan Kettering Cancer Centre
  6. Andrea Cercek; Memorial Sloan Kettering Cancer Centre
  7. Marc Gollub; Memorial Sloan Kettering Cancer Centre
Objective:
The purpose of this study was to interrogate MRI radiomics features of mesorectal fat in order to predict clinical outcomes in patient with locally advanced rectal cancers.

Materials and Methods:
In this Health Insurance Portability and Accountability Act (HIPPA) compliant, Institutional Review Board (IRB) - approved, retrospective cohort analysis of baseline MRI scans of 257 patients who underwent neoadjuvant chemoradiotherapy for locally advanced rectal cancer from 2009 to 2015, three radiologists segmented the mesorectal fat on T2WI axial MRI using Hermes Gold LX v2.3.0. After reducing the images to 32 grey levels, data was extracted from the segmented volumes and radiomic features were calculated using Computational Environment for Radiological Research (CERR). One hundred and one features were calculated in six classes. Adaptive synthetic sampling was employed to combat large class imbalance [1]. Using least absolute shrinkage and selection operator (LASSO) regression to determine important coefficients (radiomics features), a maximum of 10 features was selected. Outcome parameters from the electronic medical record included pathologic complete response (pCR), local recurrence (LR), distant recurrence (DR), clinical T- category (cT), post-treatment N category (ycN) and post-treatment T category (ycT). Predictive models were developed in Matlab using support vector machines and 5-fold cross-validation. Twenty patients were excluded, leaving a final study population of 237 patients.

Results:
The study included 237 patients (54 ± 12 years, 136 men). The average Jaccard index and dice similarity coefficient for inter-reader agreement were 0.517 and 0.669 respectively. Intra-class correlation coefficients were determined with 80% of the features demonstrating values >0.7. The area under curve (AUC), sensitivity, specificity, PPV, NPV and accuracy were; for pCR: 0.93, 80.3, 92.4, 91.8, 81.5 and 86.1, for LR: 0.88, 71.2, 91.0, 89.4, 74.8 and 80.8, for DR: 0.92, 84.4, 83.8, 84.9, 83.3 and 84.1, for (cT): 0.88, 80.4, 77.2, 77.2, 80.4 and 78.8, for ycN: 0.80, 70.6, 81.4, 80.0, 72.4, 75.8 and for ycT: 0.91, 85.4, 86.0, 85.7, 85.6 and 85.7.

Conclusion:
MRI radiomics features of mesorectal fat can help predict pathological complete response, local and distant recurrence as well as post treatment T and N categories.