Purpose: To build a radiomics model using multi-parametric pre-treatment MRI for pathological complete response (pCR) prediction with multi-center validation and compare its predictive power with two radiologists’ readings on a prospective study cohort.
Methods: This study involved an observation cohort with 143 patients for the development of radiomics model, and a prospective validation cohort of 77 patients to assess the generalization of the model. All patients receiving neoadjuvant CRT followed by total mesorectal excision (TME) with anatomical T1/T2, diffusion-weighted MRI (DWI) and dynamic contrast-enhanced (DCE) MRI collected before nCRT. A radiomics sequencer was constructed using machine-learning based approach to understand the tumor sub-regional phenotype and its association with pCR. Two radiologists with different experience (3-year and 8-year) participated the reading study. The reading study was conducted with two rounds: without and with aid of the radiomics model. In each round, the readings were repeated three times with each session at least a week apart. The consistency in terms of inter- and intra-readers agreement as well as the accuracy in assessing the pathological response without and with aids of radiomics model were evaluated.
Results: The AUC from radiomics model is 0.86 from the observational cohort and 0.84 for the validation cohort. If not providing radiomics tool, both radiologists regardless their levels of experience could only produce fair-moderate consistency in assessing the response. On contrary, with the aids of the radiomics model, both intra- and inter-reader agreements were significantly improved to good-excellent agreement. If provided with radiomics model, the accuracy in assessing the pathological response also increased significantly for both radiologists.
Conclusion: We have built a radiomics model with pre-treatment MRI for pCR prediction of LARC patients underwent nCRT and compared with two radiologists readings.
MRI, Radiation Therapy, Image Analysis