(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose:
Deep learning-based auto-segmentation (DLAS) that can substantially improve efficiency and consistency of segmentation often require large dataset, a unique training process for each tumor site and/or image modality. This can be labor intensive as multiple iterations may be required to achieve optimal accuracy. To address this problem, we developed and tested a pre-optimized module, designed to train, modify and customize segmentation models independent of tumor site and image modality, with limited training data and user input.
Methods:
The module was designed to facilitate both CT and MR images, with Z-score optimization and adaptive removal of outliers employed to achieve uniform intensity distribution for MRIs. This allowed training of MRIs independent of scanner and sequences. The training could be conducted with either U-Net or V-Net and optimized DL parameters. The training module was tested with three datasets in abdomen, 22 T1 and T2 MRIs from a 3T MR simulator, and 12 T1 MRIs from a 1.5T MR-Linac system, along with contours created based on RTOG guidelines. To prove validity of the recommended optimized parameters, multiple iterations were trained by varying the DL parameters on an Intel® Xeon® CPU E5-280 and auto-segmentation accuracy was subjected to TG 132 tolerance metrics.
Results:
The model generation with the training module was fast, requiring minimal user interaction. Reasonable accuracy was observed each time using recommended optimized parameters, e.g. batch size 16, learning rate 0.0003. For the 22 datasets, a reasonable model was trained within 30 minutes. Accuracy tests on 5 MRIs from the 3T scanner and 3 from the MR-Linac with different sequences revealed comparable performances and yielded clinically acceptable contours for majority of the organs.
Conclusion:
The training module for DLAS repeatably delivered acceptable models without a labor- and data-intensive process, a particularly useful characteristic for practical implementation of a robust auto-segmentation solution.
Funding Support, Disclosures, and Conflict of Interest: Manteia Medical Technologies
Keywords
MRI, Segmentation, Radiation Therapy
Taxonomy
IM/TH- image Segmentation: Multi-modality segmentation
Contact Email
