Author: Benjamin Haibe-Kains, Ruiyan Ni, Alexandra Rink 👨🔬
Affiliation: Department of Medical Biophysics, University of Toronto, University Health Network 🌍
Purpose: Accurate auto-segmentation for targets and organs-at-risk (OARs) using deep learning reduces the delineating time in radiotherapy. In high-dose-rate brachytherapy, specific clinical criteria need to be considered for model development, which is neglected in the existing algorithms. For High-Risk CTV (CTVHR) segmentation, we developed a multi-modal network combining MR images and clinical notes from intraoperative gynaecological examination (thereafter notes), following the ICRU89 guidelines. We also introduced novel geometrically-focused methods prioritizing regions near the CTVHR for OAR training and evaluation, aligning with the OAR dosimetric parameter (D2cm3).
Methods: 818 T2-weighted MR scans (265 cervical patients) with clinical OARs and CTVHR contours were used. Vaginal involvement information was extracted from the notes and summarized by Llama-3-8B. The multi-modal CTVHR model embedded these summaries via a text encoder into the decoder of an image segmentation model (3D U-Net). The 3D U-Net OAR model with distance-penalized loss functions prioritized errors near CTVHR during training. The proposed weighted Dice Similarity Coefficient (wDSC) emphasized the accuracy of OAR regions proximal to the target using a weighting factor based on three-dimensional dose fall-off. Predicted contours were evaluated for geometric and dosimetric accuracy, followed by clinical acceptability tests.
Results: The multi-modal CTVHR model showed significant improvement in the planar conformity index in the vaginal involvement regions compared to the image-only model (mean: 0.58 vs. 0.52, p<0.001). For OARs, wDSC moderately correlated (Pearson coefficient (r) = -0.55) with D2cm3 accuracy, outperforming standard geometric metrics in OAR models. Models with distance-penalized loss functions yielded higher wDSCs and D2cm3 accuracy compared to baseline models. Over 94% of bladder and rectum contours and about 50% of sigmoid and small bowel contours were clinically accepted.
Conclusion: We have developed models incorporating clinical information, such as distance to OARs and notes, to address practical needs in cervical brachytherapy.