Author: Asma Amjad, Renae Conlin, Eric S. Paulson, Christina M. Sarosiek 👨🔬
Affiliation: Department of Radiation Oncology, Medical College of Wisconsin 🌍
Purpose: In an effort to improve contouring accuracy for abdominal MR guided online adaptive radiotherapy (MRgOART), patient-specific deep learning-based auto-segmentation (PS-DLAS) has been proposed. In a busy clinic, this requires additional time and effort (training, contour refinement) that may not be possible. A study is conducted to explore the necessity of PS-DLAS using dosimetric and accuracy criterion.
Methods: Six patients with abdominal tumors (liver, adrenal gland, pancreas, kidney) treated with a variable imaging (3D fat/nonfat suppressed) and conventional RT (ATP: 5040cGy/28Fx) or SBRT (ATS: 3500-5000/3-5Fx) on MR-linac were included. Seven GI organs were auto-segmented using MR-based DLAS and PS-DLAS. Segmentation results were not manually edited to improve accuracy. Clinical dose-volume parameters; D0.03cc, D0.5cc, D33cc, V2500cGy, V1500cGy and Dmean, were calculated for each organ segmentation using daily adaptive plan dose distributions. Additionally, contour acceptability (ratio of clinically useful contour slices to total number of organ slices) for each organ and dice similarity coefficient (DSC) for DLAS and PS-DLAS using full organ contours within a PTV+5cm region of interest was also calculated.
Results: No discernable dosimetric differences were observed between DLAS and PS-DLAS (DSC 0.8-0.95, similar acceptability) for 3/6 patients; AvgD0.03 (cGy) of 2659/2683, 2927/2940, 3033/3116, 1937/1943 for small bowel, colon, duodenum, and stomach, AvgV1500cGy of 24.1%/22.9%, 9.2/9.8% for both kidneys and Dmean of 284/282 cGy for liver. Improved dosimetric outcomes were observed using PS-DLAS in 2 patients (low DSC, high acceptability of PS-DLAS), one identified as an imaging outlier (poor contrast) and the other as an anatomical outlier.
Conclusion: A segmentation model trained with anatomical and imaging variations is suitable for clinical ATP/ATS workflows. PS-DLAS should only be introduced for patients with abnormal anatomy or poor imaging (e.g., large bowel changes and surgical procedures) thus saving extra time and effort otherwise required to employ PS-DLAS for all patients.