Patient Fine-Tuning of NnUnet for MRI Organ Autosegmentation
J Hogan1, D Yang2*, B Baumann3, Y Duan4, C Zhao4, (1) Washington University In St. Louis School of Medicine, St. Louis, MO, (2) Duke University, Department of Radiation Oncology, Chapel Hill, NC, (3) Washington University In St. Louis School Of Medicine, Department of Radiation Oncology, St. Louis, MO, (4) University Of Missouri, EECS Department, Columbia, MO
Presentations
TU-D930-IePD-F9-1 (Tuesday, 7/12/2022) 9:30 AM - 10:00 AM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 9
Purpose: Deep learning (DL) auto-contouring has become increasingly prevalent in radiation oncology treatment planning. However, DL methods often require an in-depth knowledge of coding, limiting clinical applications. Furthermore, anatomical variation between patients can negatively affect DL segmentation methods. To our knowledge, a fully-automated patient-specific DL segmentation model has not been published. Our study aims to (1) provide an auto-contouring solution that increases accessibility, and (2) use patient-specific algorithm fine-tuning to improve model specificity.
Methods: nnUNet is a PyTorch-based tool that has high performance on diverse image segmentation tasks. In this study, nnUNet was trained and validated on MRI scans from 110 patients to segment 5 abdominal organs and tested on scans from 10 patients. Validation results for six parameter settings were explored, and the model was selected based on high DICE coefficients (DSC) and low model complexity. The model was further fine-tuned to be patient-specific using an individual patient’s initial treatment planning data and evaluated using the respective patient’s fractional data.
Results: None of the six parameter settings had significantly different DSC from the others (Kruskal-Wallis p > 0.05 in all cases). The simplest model, with 8 base features (compared to 32 in the standard configuration), had the lowest computational burden and was thus selected. The average DSC of the generalized network were 94.6±3.7, 89.2±10.6, 89±3, 65.4±14.4, and 87.4±3.6 for the liver, kidneys, stomach, duodenum, and bowel respectively.
Conclusion: Our model provides improved computational efficiency, increased accessibility, and the ability to segment full-resolution MRI images, with similar performance when compared to a previous DL model which can only segment on down-sampled MRIs. Our results also show the potential for patient-specific fine-tuning to further increase contouring accuracy. These results aim to encourage more widespread adoption and subsequent improvement of auto-contouring methods in the clinical space.
Funding Support, Disclosures, and Conflict of Interest: BCB reports reimbursement for the following: Varian consulting work and research funding, Boston Scientific consulting work, Regeneron/Sanofi medical advisory panel, and Galera medical advisory panel (all outside scope of submitted work).
Keywords
Taxonomy
Not Applicable / None Entered.
Contact Email