Improving Deep Learning Auto-Segmentation Using An Adaptive Spatial Resolution Approach
A Amjad1*, J Xu2, D Thill2, M Awan1, M Shukla1, W Hall1, B Erickson1, X Li1, (1) Medical College of Wisconsin, Milwaukee, WI, (2) Elekta Inc., MO, USA
Presentations
TU-IePD-TRACK 4-4 (Tuesday, 7/27/2021) 3:00 PM - 3:30 PM [Eastern Time (GMT-4)]
Purpose: Deep learning-based auto-segmentation (DLAS), although has been shown to substantially improve organ delineation efficiency, performs generally poor for organs with small/narrow size (e.g., inner ears) or complex shape (e.g., bowels). To overcome this problem, we introduce an adaptive spatial resolution approach (ASRA) for DLAS.
Methods: The framework of ASRA uses a fine spatial resolution in a subregion that contains small/narrow and/or complex anatomy while using default or coarse resolutions in other regions to balance the accuracy and efficiency of DLAS. We implemented ASRA in 3DResUnet and trained the DLAS models for selected organs in head and neck (HN) based on CT by using a fine spatial resolution of 1×1×2 mm³ inside the subregions, and 2×2×2 mm³ outside it. To overcome a common issue that a portion of a large organ (e.g., small bowel) far away from the target may not delineated or may be missing in the training dataset, we extended existing contours via 8-fold repetition of last contoured slice using 3D elastic transformation, thus stimulating a larger field-of-view. DLAS models were trained with 47 datasets for HN and 58 for abdomen. Performance of the obtained models with and without ASRA was compared.
Results: DLAS accuracy was substantially improved with the use of ASRA compared to that without ASRA, with maximum improvement observed for small/narrow organs, e.g., eye lens and optic nerves with dice similarity coefficient (DSC) of 0.74 – 0.8 vs. 0.25 – 0.45. For complex structures, e.g., small and large bowels, DSC/mean distance to agreement (MDA) of 0.84/3.8 mm and 0.87/2.37 mm vs 0.71/7.9 mm and 0.7/14 mm, were observed with 3D extended volume respectively.
Conclusion: The proposed adaptive spatial resolution approach can substantially improve accuracy of deep learning-based auto-segmentation, particularly for organs with small/narrow sizes or complex shapes.
Funding Support, Disclosures, and Conflict of Interest: The work was partially supported by the Medical College of Wisconsin (MCW) Cancer Center and Froedtert Hospital Foundation, the MCW Fotsch Foundation, Elekta AB, and the National Cancer Institute of the National Institutes of Health under award number R01CA247960.
ePosters
Keywords
Taxonomy
Contact Email