ESTRO 2023 - Abstract Book

S94

Saturday 13 May

ESTRO 2023

TL25 outperformed the RTmodel when testing on S-N cases [mean vDSC of 0.88 vs 0.81 for bladder, 0.88 vs 0.74 for rectum, 0.69 vs 0.57 for sigmoid, 0.65 vs 0.48 for small bowel, and 0.72 vs 0.54 HR-CTV]. The Mixed model had similar performance compared with both RTmodel for R&T cases and TL25 for S-N cases, indicating that mixing different applicators does not reduce the model performance (improved mean DSC of 0.07 for S-N small bowel, diff < 0.02 for others). TL significantly improved the model performance with only 5 finetuning cases (mean vDSC increased 0.03~0.14) but reached a plateau with ≧ 10 finetuning cases (mean vDSC diff < 0.03). The training-from-scratch models performed worse than the TL models. The training times of RT/Mixed model and one TL model were 14 hours and 2.5 hours. The prediction time per image was 12 s.

Conclusion We have successfully demonstrated that (1) DL model has the ability to handle segmentation with various applicators; (2) TL can achieve similar results to Mixed model with limited fine-tuning data and highly reduced computational costs. This study shows the potential of TL when applying our model to different institutions in the future. OC-0132 Near-to-target-aware OARs segmentation in cervical HDR brachytherapy via deep learning R. Ni 1 , B. Haibe-Kains 1,2 , A. Rink 1,2,3,4 1 University of Toronto, Department of Medical Biophysics, Toronto, Canada; 2 University Health Network, Princess Margaret Cancer Center, Toronto, Canada; 3 University of Toronto, Department of Radiation Oncology, Toronto, Canada; 4 University Health Network, TECHNA Institute, Toronto, Canada Purpose or Objective Deep learning (DL) has been used to automate and speed up the time-consuming manual contouring step for organs and targets in radiotherapy. In cervical high dose-rate brachytherapy (HDR-BT), the delineation of OARs far from the target is not considered during treatment planning because treatment planning is driven by OAR dose to 2 cubic centimeters (D2cm ³ ) closest to the target. Geometrics in distal OAR segmentation does not necessarily translate to dosimetric or clinical impact. A novel loss function was developed in this study to guide the network’s attention and ensure highly accurate OAR segmentation close to the applicator which will increase the clinical relevance and acceptability of the model’s predictions. Materials and Methods A dataset of 130 T2-weighted MR images with clinically used contours was built from 39 cervical cancer patients undergoing HDR-BT. Four OARs (bladder, rectum, sigmoid, and small bowel) were segmented. The distance-penalized (DP) map was calculated by (1) generating a distance transform map with HR-CTV as the center; (2) applying inverse square law and normalizing by each organ (range: [0,1]); (3) adding one to the map and weighting it by values, as shown in Figure 1. The generated map D was used to penalize the training errors. The training goal is to minimize the penalized multi-class combo loss L in Equation (1), where L_DPCE and L_DPDice are the DP cross-entropy loss and DP Dice loss; N and C are the numbers of voxels and classes. The network was trained on 119 cases (35 patients) with 5-fold cross-validation via 3D U-Net. Combo loss without DP ( L_CE + L_Dice ) was used as the baseline for comparison. The segmentation performance was evaluated by the volumetric Dice Similarity Coefficient (DSC) and a newly proposed weighted DSC (wDSC) as shown in Equation (2), where A , B and dist are the ground truth, prediction and penalty term for each class, respectively.