Author: Austin Castelo, Xinru Chen, Caroline Chung, Laurence Edward Court, Jaganathan A Parameshwaran, Zhan Xu, Jinzhong Yang, Yao Zhao 👨🔬
Affiliation: The University of Texas MD Anderson Cancer Center, Department of Radiation Physics, The University of Texas MD Anderson Cancer Center, Department of Imaging Physics, The University of Texas MD Anderson Cancer Center 🌍
Purpose:
To develop a deep learning-based segmentation model to automatically delineate tumors from full-body PET/CT images.
Methods:
PET/CT image pairs of 91 patients were collected for this study, with primary tumor sites in lungs (34), liver (28), head and neck (16), female pelvis (11), and male pelvis (2). Two radiologists manually contoured the tumors, which were later reviewed and edited by two other radiologists for consistency. For each patient, number of tumors varies from 1 to 21, and the tumor size varied from 0.6 cc to 792.3 cc. A 2-stage framework was developed for tumor auto-segmentation. In stage 1, 25 CT images were randomly selected from the dataset to train an nnU-Net model to auto-segment brain, heart, bilateral kidneys, and bladder. The segmented masks of these organs were then applied to PET images to exclude high standardized uptake values in these organs. In stage 2, the masked PET and CT images were fed into a multi-modal segmentation framework[1] based on nnU-Net to delineate tumors automatically. 20 patients were reserved for test and 71 were used to train the multi-modal segmentation model with a five-fold cross-validation. Model performance was evaluated using sensitivity and positive prediction value (PPV) for tumor detection and Dice Similarity coefficient (DSC) for segmentation accuracy.
Results:
In cross-validation, the tumor detection sensitivity was 0.90, PPV was 0.84, and detected tumors had an average DSC of 0.80. In testing, sensitivity was 0.84, PPV was 0.73, and DSC was 0.76. Performance varied in different tumor sites due to the lack of sufficient training/testing data. Lung and liver have the most data and showed better performance than other sites (Lung: sensitivity 0.92, PPV 0.77, DSC 0.77; Liver: sensitivity 1.0, PPV 0.76, DSC 0.79 for testing dataset).
Conclusion:
The auto-segmentation framework has been demonstrated effective in full-body tumor delineation from PET/CT images.