Cross-Treatment Knowledge Transfer: Generalizability of a Deep Learning Autosegmentation Model Trained on External Beam Contours for Brachytherapy
Abstract
Purpose
To evaluate the generalizability of a deep learning (DL)-model, trained exclusively with external beam radiotherapy (EBRT) organs‑at‑risk (OARs) contours, for autocontouring in cervical cancer brachytherapy (CC-BT).
Methods
Manual OAR contours (bladder, rectum, sigmoid, bowel) from EBRT CTs (without CC-BT applicator, n=31 patients) were used to train/validate an nnU-Net-based autocontouring model. Our clinical CC-BT workflow includes 2 insertions delivering 2 HDR fractions (Fx) each: Fxs1&3 are MR-planned and Fxs2&4 CT-verified. At-BT CTs from 42 cases (with applicator; n=24 Fx2, n=18 Fx4) having manual OAR contours (ground truth, GT) were used for DL-model testing. Geometric evaluation of GT vs predicted contours included Dice Coefficient (DC), volume similarity, Hausdorff distance, and precision; DC reported here. Clinically delivered MR plans from Fxs1&3 were projected on the predicted contours for Fxs2&4, respectively, and dosimetric analysis performed to determine OAR D2cc differences (ΔD2cc; GT-predicted).
Results
Training and prediction times were 1.6 hours and 4.4 sec/case, respectively (NVIDIA H100 80GB GPU). For bladder, the mean±SD DC was 0.69±0.19, rectum 0.68±0.14, sigmoid 0.49±0.22, bowel 0.22±0.15; similar trend seen for other geometric metrics. The moderate DC values reflect the impact of domain shift, small training dataset, and applicator-induced anatomic displacement. Variations in EBRT vs BT practices (bladder full vs half-full to empty, bowel bag vs loops contoured) were also observed to influence evaluated metrics. However, dosimetric analysis showed low mean ΔD2cc (Gy) indicative of geometric inaccuracies being located outside of high-dose regions; for bladder -0.1±0.4, rectum -1.7±2.5, and sigmoid 0.1±0.5, but higher for bowel -6.7±9.6. DC did not differ significantly between BT Fxs (p>0.5).
Conclusion
The proposed autocontouring model, trained on EBRT contours, provides reasonable performance for use in brachytherapy (except for bowel), supporting knowledge transfer across techniques. This generalizable approach may be considered in data-limited settings until adequate data is acquired for developing a robust task-specific model.