Boundary-Aware Mamba Improves Cervical Radiotherapy Auto-Contouring

Radiotherapy planning for cervical cancer depends on accurate contours of target volumes and organs at risk, but manual delineation is slow and varies with reader experience. That bottleneck is sharper in settings where radiation oncologists are scarce and treatment queues are long. This paper presents BAT-RM, a deployed auto-contouring system built around a boundary-aware transformer and a region-aware multi-directional Mamba module, then evaluates it from dataset curation through clinical use.

The work is stronger than a model-only segmentation paper because it connects architecture design to workflow evidence. The authors report multi-institutional data collection, inter-rater quality assurance, external validation, a web interface compatible with Varian, RayStation, and Monaco, a prospective reader study with 13 radiation oncologists, and deployment at a partner hospital. The central claim is that boundary-sensitive long-range modeling can improve both contour accuracy and operational throughput.

Core Contribution

BAT-RM targets a specific failure mode in radiotherapy segmentation: anatomically plausible volumes can still have clinically unacceptable boundary errors. The paper therefore gives boundary modeling first-class status rather than treating it as an afterthought of volumetric Dice optimization. Its novelty is the combination of three components: Sobel-gated boundary attention for edge-sensitive features, region-aware multi-directional Mamba blocks for linear-time long-range context, and a boundary-skeleton-guided fusion gate that merges regional and boundary streams.

That design choice fits the task. Cervical radiotherapy contours include large target volumes such as GTV and CTV, but also organs at risk whose small boundary deviations can change dose planning. A method that improves only average overlap would be less convincing here; the paper emphasizes Dice, Average Surface Distance, and HD95, which better expose boundary failures.

Technical Approach

The architecture couples a Boundary-Aware Transformer path with a Region-Aware Mamba path. The Mamba branch is used to model long-range spatial dependency with linear-time scaling, avoiding the quadratic cost of full spatial self-attention. The boundary path uses Sobel-derived signals to bias the network toward edges, and the final decoder fuses these sources using a learned spatial-channel attention gate guided by a predicted boundary skeleton.

Figure 7 shows the Boundary-Region Attention Fusion module: region-aware Mamba features are upsampled and channel-projected, aligned with boundary-aware transformer features, passed through a learnable attention gate, and refined using the boundary skeleton. The mechanism is clinically sensible because it asks the network to reconcile interior region evidence with explicit contour evidence before producing decoder features.

The paper also spends visible effort on data quality. Its dataset characterization covers class occurrence frequency, foreground-background voxel imbalance, voxel spacing, and voxel percentage distributions. Inter-rater reliability is reported with pairwise Jaccard similarity of 0.961 ± 0.024, which supports the claim that the labels were consistent enough to train and evaluate against.

Results and Analysis

Against four segmentation baselines, BAT-RM is reported as the top model across seven anatomical classes. The figures show paired violin-box comparisons for Dice, ASD, and HD95, with Wilcoxon testing and Bonferroni-Holm correction. The clearest message is not just higher central tendency; BAT-RM also shows narrower error distributions, especially on boundary metrics. For radiotherapy, that matters more than a small mean Dice gain because rare large contour errors can force manual correction.

Figure 8 summarizes Dice across GTV, CTV, bladder, rectum, small bowel, and femoral head classes. The paired design is useful: it compares models on the same cases rather than treating cohort-level distributions as independent samples. The surface-distance figures reinforce the same pattern, with BAT-RM showing lower ASD and fewer HD95 outliers, while baseline models display heavy-tailed failures above 10 mm for small bowel, rectum, and femoral head.

The prospective reader study is the most practice-facing evidence. With AI assistance, junior oncologists improve from 0.899 IoU to 0.965 IoU, approaching senior-level accuracy, while contouring time falls by more than 80%. The paper also reports fewer expert consultations and improved inter-reader consistency. Those numbers support a concrete workflow interpretation: BAT-RM is less a replacement for clinicians than a quality-assurance and acceleration layer for less experienced readers.

Caveats

The evidence is broad for one disease site but still narrow in scope. The system is evaluated for cervical cancer radiotherapy, not a general oncology contouring setting. The abstract and figures support strong claims about segmentation quality and workflow speed, but the available material does not show long-term outcome data such as replanning rates, toxicity, local control, or survival. Deployment reduced waits from days to hours at a partner hospital, which is meaningful operational evidence, but the article should treat that as site-specific until replicated across more hospitals and scanner/protocol mixes.