Author: Asma Amjad, Renae Conlin, Eric S. Paulson, Christina M. Sarosiek 👨🔬
Affiliation: Department of Radiation Oncology, Medical College of Wisconsin 🌍
Purpose:
MR-guided adaptive radiation therapy (MRgART) is transforming clinical workflows, requiring fast, accurate organs-at-risk (OARs) contouring. While deep learning auto-segmentation (DLAS) offers a promising solution, the variability and lack of standardization in MR imaging protocols challenge accuracy, especially compared to CT-based DLAS. This study examines training challenges for abdominal DLAS models using MR-Linac (MRL) images.
Methods:
Over five years, thirteen abdomen DLAS models were trained for MRgART using algorithms A and B across eight MR imaging protocols. Twelve models were trained with algorithm A, incorporating 1-5 protocols and various preprocessing. Three models used algorithm B, incorporating 1-3 protocols without preprocessing. For CT, three models used algorithm A with two protocols and an increasing dataset. Performance was evaluated using Dice similarity coefficient (DSC) and mean distance to agreement (MDA).
Results:
Two MRL models were clinically acceptable, achieving accuracy suitable for MRgART workflows. One (DLAS A) used algorithm A with 79 images from one protocol (btFE(f) with online inhomogeneity correction). The other (DLAS B) used algorithm B with 30 images from two protocols (15 each of btFE(f) and tFE(f)). Algorithm A excelled with one MR protocol, while algorithm B did well with multiple protocols and unseen data. Poor image quality and excess protocols resulted in insufficient intensity normalization and lower performance in other models. The best CT DLAS model outperformed MRL models for GI structure accuracy, with mean DSC (0.87, 0.82, 0.79) and MDA (2.28, 5.99, 9.99) across the small bowel, colon, duodenum, and stomach for CT, DLAS A, and DLAS B, respectively.
Conclusion:
CT-based DLAS training is straightforward, with a single protocol and additional data improving accuracy. In MRL DLAS, training strategies, notably algorithm and protocol selection, significantly impact performance, while preprocessing has little effect. However, abdominal MRL variability prevented MRL DLAS from matching the best CT DLAS accuracy.