External validation of a deep learning CBCT auto-segmentation model for lung tumor and esophagus in online adaptive radiotherapy

Source article: External evaluation of a deep learning-based algorithm for automated lung tumor and esophagus segmentation on CBCT

Online adaptive radiotherapy (ART) requires fast and accurate delineation of targets and organs at risk on cone-beam computed tomography (CBCT). Although deep learning (DL)-based CBCT auto-segmentation methods have been proposed, their clinical generalizability remains insufficiently validated. This study aimed to externally evaluate a pretrained DL model for lung tumor and esophagus segmentation on CBCT. Originally trained using C-arm linac CBCT images, the model was directly applied to an independent cohort of…

By the Truvace deskOctober 14, 20262 min read1 sourcePeer-reviewedTRV-2026-1365· cite & verify

External evaluation of a deep learning-based algorithm for automated lung tumor and esophagus segmentation on CBCT — Metastatic follicular thyroid carcinoma - CT scan - Case 264 (8559838032) by Yale Rosen from USA. CC BY-SA 2.0 · https://creativecommons.org/licenses/by-sa/2.0

Trace impact readingContested

P 66

Both sides are scored from claims and sources, not community votes.

G 63

In brief

Researchers tested a pretrained deep learning model, originally trained on C-arm linac CBCT images, on 12 lung cancer patients treated at another institution on an O-ring linac. Each patient had 5 CBCT scans. The model's tumor and esophagus contours were compared with rigid propagation of planning CT contours and with reference contours.

For tumor segmentation the model clearly beat the rigid baseline, but its performance and some dose measures fell short of what was achieved on its original test data. This matters for online adaptive radiotherapy, where contouring must be both fast and reliable, and it shows that results from a development site may not carry over. Which adaptations would close the gap is not addressed.

Main points

Tumor Dice rose from 0.66 with rigid propagation to 0.76 with the model, a significant improvement.
On the original test dataset the model had reached a mean Dice of 0.84 for tumors; the drop to 0.76 was not statistically significant.
Esophagus Dice was 0.72 for the model versus 0.70 for rigid propagation, which the authors call comparable.
Esophagus Dice fell significantly from 0.79 on previously reported results to 0.72 in this evaluation.
Tumor D98% differed significantly between model and reference contours (50.69 vs. 53.58 Gy), while esophagus V32Gy was similar (0.06 vs. 0.06 cc).

The gain

On an independent cohort from a different institution, the pretrained model segmented lung tumors on CBCT more accurately than rigid contour propagation, raising mean Dice from 0.66 to 0.76.

The problem

Applied outside its original development environment, the same model lost accuracy, with esophagus Dice falling from 0.79 to 0.72 and tumor dose metrics differing from reference contours.

The rundown

The external cohort differed from the training setting in institution and imaging system: images came from an O-ring linac rather than the C-arm linac used for training. The model was applied directly without retraining, and rigid propagation of planning CT contours served as a minimal baseline.

Esophagus evaluation used 26 scans from 6 patients and focused on peri-target contour rings. In that region the model matched rigid propagation geometrically and gave similar esophagus V32Gy values to the reference contours.

The authors conclude that CBCT auto-segmentation may degrade under domain shifts across patient populations, clinical workflows, and imaging systems, and that external evaluation beyond the original development environment is necessary.

What this doesn’t fix

The evaluation used a small cohort of 12 lung cancer patients from a single external institution, with esophagus results based on only 6 patients, so the findings may not generalize widely.