Room: AAPM ePoster Library
Purpose: purpose of this study is to demonstrate the clinical utility of a robust automatic CT segmentation software for radiation treatment planning using deep-learning algorithms.
Methods: neural networks are used for the method: for CT image classification and for the segmentation tumor and organs-at-risk. The image classification network is based on 4 different feature extraction schemes that achieve accurate object identification of CT images. The results of the classification include the position of each CT image slice in the patient body. Subsequently, the tumor and OARs are segmented according to the position of that image slice. The segmentation network is based on the U-Net that is combined with special residual modules. A total of 350 patients data are used as training data in this study. Manually segmented CT image datasets by experienced radiologists are used as the ground-truth. Dice Similarity Coefficients (DSCs) of targets and OARs are calculated. A user graphical interface (GUI), called DeepViewer, is developed to integrate the validated CNNs with DICOM processing features. The clinical utility of the software is evaluated as part of the very busy workflow in more than 30 radiation oncology departments that, together, treats more than 3000 patients per day.
Results: chest and abdomen CT scan patients were tested, the average DSCs of organs are found to be 0.97 (right lung), 0.96 (left lung), 0.92 (heart), 0.86 (spinal cord), 0.76 (esophagus), 0.96 (spleen), 0.96 (liver), 0.95 (left kidney), 0.90 (stomach), 0.87 (gall bladder), 0.80 (pancreas), and 0.61 (duodenum). The average DSCs of tumor clinic target volume (CTV) for 30 test patients are found to be 0.81 (esophagus cancer), 0.82 (breast cancer), 0.80 (cervical cancer) and 0.82 (nasopharynx cancer).
Conclusion: DeepViewer software that incorporates DL-based automatic tumor and OAR segmentation models has been shown to improve the workflow for routine radiation treatment planning.