Z. R. Li1, Q. Zhou1, and X. Qi2; 1Department of Research Algorithms, Manteia Medical Technologies Co., Milwaukee, WI, 2UCLA, Los Angeles, CA
Purpose/Objective(s): Radiomics demonstrates great promises and has been used to predict radiation treatment response for many cancer sites including locally advanced rectal cancer (LARC). Traditional radiomics based prediction is generally based on quantitative features derived from 3D volumes in CT/MR images. For the hollow anatomic structure of the organs, such as rectum/bladder, the 3D contour-based features may contain extraneous irrelevant features, resulting in inferior response prediction performance. We aimed to develop a novel radiomic framework based on unfolded images for the hollow organs, with the hope to better capture the distinctive relevant features for accurate prediction towards personalized treatment.Materials/
Methods: The innovation of the proposed radiomic framework is to extract imaging features from virtual rectum-unfolding in multi-parametric MR images. The imaging unfolding process consists of three steps: starting from the centroid of rectal delineation structures as the initial point, a projection light source was created to emit outward. The values of each pixel grid point traversed by the light path were accumulated. Next, the projection lines were sorted based on the angle of the light path and unfolded to obtain the projection lines of each layer. Finally, all unfolded projection lines were stacked in the original Z-direction order to obtain the projection images. The unfolding algorithm, as well as the proposed prediction framework, was evaluated using a cohort of 42 LARC patients who underwent neo-adjuvant radiation treatment, followed by total mesorectal excision surgery. Patient-specific response was assessed by post-operative pathology report. Imaging features extracted from p<span>re-treatment MRI. The U-test was performed to select features. The final p</span>rediction models, using the random forest algorithm, were constructed using the proposed framework and the traditional contour-basedradiomic framework. The performance of the models was evaluated using the areaofthecurve (AUC). Results: Post-RT pathology scores revealed 22 responders (complete response and partial response) and 20 non-responders (stable disease and progressive disease).Using leave one out cross validation, we calculated the AUC using each patient as the predicted result of the test set. The AUC of the proposed framework achieved 0.73,whereas the AUC using traditional contour-based methods was 0.63. Conclusion: The proposed framework achieved better prediction accuracy to stratify post-RT response based on pre-treatment MRIs for LARCs. The imaging features derived from the unfolded images better capturedrelevant underlying tumor characteristics, resulting in an improvedprediction accuracy and robustness compared to traditional volumetric contour-based radiomic model. The proposed framework may aid clinical decision support for development of personalized management towards optimal treatment response.
