A Generalizable Target Contour Validation Method Using Radiomics Based Regression Analysis
Z Lin1*, C Chen3, Z Fei3, Q Zhou2, z li2 (1) Xiamen University,(2) Manteia Medical Technologies, Xiamen,(3)CN,Department Of Radiotherapy, Fujian Cancer Hospital, Fujian Medical Universi
Presentations
PO-GePV-M-251 (Sunday, 7/25/2021) [Eastern Time (GMT-4)]
Purpose: To analyze the changing trend of radiomics features of the targeted volumes, especially those with fuzzy edges, on the boundary area in CT images of nasopharyngeal carcinoma patients, and to explore this approach as a method to evaluate the quality of automatic or manual delineation of GTVs and organ at risks (OARs) with low contrast boundaries.
Methods: Contours of GTVs and OARs in CT images of nasopharyngeal carcinoma patients were expanded and retracted by 1 mm, 2 mm and 3 mm respectively. Next, extract the radiomics features of the original image, the expanded images and the shrunk images. Then, the difference of adjacent features is calculated to get the feature differences. Regression model is built on features extracted using decision tree algorithm. The error of the model in the test set is calculated and analyzed. Performance was assessed on a separate test dataset.
Results: We conducted sufficient experiments on 2 datasets of nasopharyngeal cancer patients. One set of the data was used as the ground truth for model training, and the other data set was used for model testing. We used the mean square error as the evaluation index of the model, and the error is 0.02. We compared the results of the model prediction to doctor's judgment, and found that our model can distinguish and gives evaluation on these situations very well: incomplete delineation of the area, delineation area is too large, and delineation is of good quality.
Conclusion: We found that there are changing trend of radiomics features in the boundary area of targeted volumes with fuzzy edges in CT images of nasopharyngeal carcinoma patients. By exploring this characteristic change using regression models, we can evaluate the quality of the target area delineation automatically to accelerate quality assurance procedure in clinical practice.
ePosters
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email