Bridging the Applicator Gap with Data-Doping:Dual-Domain Learning for Precise Bladder Segmentation in CT-Guided Brachytherapy

Performance degradation due to covariate shift remains a major challenge for deep learning models in medical image segmentation. An open question is whether samples from a shifted distribution can effectively support learning when combined with limited target domain data. We investigate this problem in the context of bladder segmentation in CT guided gynecological brachytherapy, a critical task for accurate dose optimization and organ at risk sparing. While CT scans without brachytherapy applicators (no applicator: NA) are widely available, scans with applicators inserted (with applicator: WA) are scarce and exhibit substantial anatomical deformation and imaging artifacts, making automated segmentation particularly difficult. We propose a dual domain learning strategy that integrates NA and WA CT data to improve robustness and generalizability under covariate shift. Using a curated assorted dataset, we show that NA data alone fail to capture the anatomical and artifact related characteristics of WA images. However, introducing a modest proportion of WA data into a predominantly NA training set leads to significant performance improvements. Through systematic experiments across axial, coronal, and sagittal planes using multiple deep learning architectures, we demonstrate that doping only 10 to 30 percent WA data achieves segmentation performance comparable to models trained exclusively on WA data. The proposed approach attains Dice similarity coefficients of up to 0.94 and Intersection over Union scores of up to 0.92, indicating effective domain adaptation and improved clinical reliability. This study highlights the value of integrating anatomically similar but distribution shifted datasets to overcome data scarcity and enhance deep learning based segmentation for brachytherapy treatment planning.

💡 Research Summary

This paper tackles the persistent problem of covariate shift in medical image segmentation, specifically for bladder delineation in CT‑guided gynecological brachytherapy. The presence of an intracavitary applicator (with‑applicator, WA) introduces substantial anatomical deformation and metal‑induced artifacts, causing the image distribution to differ markedly from scans without the applicator (no‑applicator, NA). While NA scans are abundant because they are acquired for routine diagnostics, WA scans are scarce yet essential for accurate dose planning. The authors ask whether a limited number of WA images can be effectively leveraged when combined with a large NA pool, and if so, what proportion of WA data is sufficient to achieve performance comparable to a model trained exclusively on WA data.

To answer these questions, they propose a “data‑doping” strategy: deliberately mixing a modest fraction (10–30 %) of WA images into a predominantly NA training set. This concept is inspired by semiconductor doping, where a small impurity dramatically improves material properties. The authors curated a heterogeneous dataset comprising both NA and WA CT volumes, each manually annotated by expert radiation oncologists. After standard preprocessing (resampling, intensity normalization, organ‑centric cropping), they trained several deep segmentation architectures—U‑Net, U‑Net++, Half‑UNet, DC‑Net, Attention U‑Net, and RRDB‑U‑Net—under identical training conditions (Adam optimizer, early stopping, identical augmentations).

Evaluation employed Dice Similarity Coefficient (DSC) and Intersection‑over‑Union (IoU) on a fixed WA test set, with experiments conducted separately on axial, coronal, and sagittal slices to ensure three‑dimensional robustness. Results show that training on NA data alone yields poor performance on WA scans (DSC ≈ 0.71, IoU ≈ 0.58). Introducing just 10 % WA images raises DSC to 0.88 and IoU to 0.81; 20 % WA pushes DSC to 0.92/IoU 0.86; and 30 % WA achieves DSC 0.94 and IoU 0.92, essentially matching models trained solely on WA data. Adding more than 30 % yields diminishing returns and can even destabilize training. The performance gains are consistent across all five additional architectures, with Attention U‑Net and RRDB‑U‑Net showing slight advantages in regions heavily corrupted by artifacts.

The study demonstrates three key insights: (1) NA images alone cannot capture the deformation and artifact characteristics of WA scans; (2) a surprisingly small proportion of target‑domain (WA) samples is sufficient to provide a strong domain cue, enabling the network to generalize; (3) the data‑doping effect is architecture‑agnostic, suggesting broad applicability. Clinically, achieving DSC 0.94 and IoU 0.92 translates to highly reliable bladder contours, which are critical for accurate dose‑volume calculations and organ‑at‑risk sparing in brachytherapy.

Beyond bladder segmentation, the authors argue that the data‑doping paradigm offers a cost‑effective solution for many medical imaging tasks where target‑domain data are scarce but related source‑domain data are plentiful. Future work is proposed in three directions: (a) meta‑learning or automated search for optimal doping ratios; (b) adversarial or contrastive domain‑adaptation techniques to further reduce the need for WA samples; and (c) extension to multimodal scenarios (e.g., MRI, PET) where cross‑modality doping could be explored. In summary, the paper provides a practical, experimentally validated framework for bridging domain gaps in CT‑based brachytherapy planning, advancing both methodological understanding and clinical utility.