In:
Cancer Research, American Association for Cancer Research (AACR), Vol. 83, No. 7_Supplement ( 2023-04-04), p. 5374-5374
Abstract:
Introduction As total neoadjuvant therapy for locally advanced rectal cancer (LARC) emerged, the possibility of skipping radiotherapy for poorly responsive patients arose. Machine-learning algorithms have focused on the radiopathologic features of tumor segmentation in order to predict responsiveness to radiotherapy. However, in addition to the tumor itself, there are several factors related to responsiveness, such as vasculature affecting hypoxia or MRI-detected extramural vascular invasion status. We aimed to predict poor responders using pretreatment rectal MRI images without segmentation and to identify which factors mainly contribute to the prediction algorithm. Methods Between Jan 1, 2000, and Dec 30, 2020, 689 consecutive patients were retrospectively included in two hospitals. Poor responders were defined by tumor regression grades (TRG) 2 and 3 that were determined through surgical resection. The ResNet-50 model was trained to predict poor responders from pretreatment rectal MRIs (T2-weighted axial, sagittal, and coronal images). We adopted a tenfold cross-validation for training and testing the model and used Gradient-weighted Class Activation Mapping (Grad-CAM) to highlight the important regions in the MRI scans that help predict poor responders. Results The number in each group of TRG was 108 (15.7%), 250 (36.3%), 265 (38.5%), and 66 (9.6%) for TRG0, TRG1, TRG2, and TRG3, respectively. There were 618 patients in the training cohort and 71 patients in the validation cohort. In the training and validation cohort, the accuracy for the prediction of poor responders was 85.6% (area under the curve (AUC) 0.856 [95% CI 0.761-0.950]) and 70.2% (AUC 0.703 [0.682-0.724] ), although without segmentation. Our prediction model achieved a sensitivity of 0.724 (95% CI 0.700-0.748), a specificity of 0.684 (0.658-0.710), a positive predictive value of 0.697 (0.656-0.737), and a negative predictive value of 0.708 (0.666-0.751) in the validation cohort. Grad-CAM showed that the most important part of the accurately predicted images to contribute to the prediction was not the tumor (7/355, 1.9%) but the pelvic vasculature (353/355, 99.4%), including iliac vessels, femoral vessels, and presacral plexus, and followed by the mesorectum (38/355, 10.7%). Conclusion The pelvic vasculature contributes more to predicting poor responders to radiotherapy than the tumor itself. Therefore, when creating a prediction model for responsiveness to radiotherapy in LARC, this should be considered. Citation Format: Rumi Shin, Byunho Jo, Inyeop Jang, Cheong-Il Shin, Jin Sun Choi, Seung-Yong Jeong, Seung Chul Heo, Ji Won Park, Min Jung Kim, Tae Hyun Hwang. Why only focus on the tumor?: The crucial role of the extra-tumor environment to predict poor responders for locally advanced rectal cancer. [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 5374.
Type of Medium:
Online Resource
ISSN:
1538-7445
DOI:
10.1158/1538-7445.AM2023-5374
Language:
English
Publisher:
American Association for Cancer Research (AACR)
Publication Date:
2023
detail.hit.zdb_id:
2036785-5
detail.hit.zdb_id:
1432-1
detail.hit.zdb_id:
410466-3
Permalink
