⚡ electrical engineering

BAT-RM: A Boundary-Aware Transformer with Region-Aware Multi-Directional Mamba for Clinically Deployed Cervical Cancer Radiotherapy Auto-Contouring

The paper presents BAT-RM, a clinically deployed hybrid architecture combining a Boundary-Aware Transformer with Region-Aware Multi-Directional Mamba for cervical cancer radiotherapy auto-contouring, which achieves superior accuracy, reduces contouring time by over 80%, and enables same-day treatment initiation by bridging the gap between advanced AI research and real-world clinical deployment.

Original authors: Istiak Ahmed, Kazi Shahriar Sanjid, Galib Ahmed, Md. Tanzim Hossain, Md. Anwarul Islam, Shahrukh Khan, Md. Ashrif Rahman Arian, Md. Nishan Khan, Md. Misbah Khan, S M Hasibul Hoque, Rahnuma Shahrin Ris

Published 2026-07-15

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Istiak Ahmed, Kazi Shahriar Sanjid, Galib Ahmed, Md. Tanzim Hossain, Md. Anwarul Islam, Shahrukh Khan, Md. Ashrif Rahman Arian, Md. Nishan Khan, Md. Misbah Khan, S M Hasibul Hoque, Rahnuma Shahrin Rista, Md. Jobairul Islam, Sheikh Anisul Haque, Md Arifur Rahman, Syed Md. Akram Hussain, Syeda Nashra, Sayeed Shafayet Chowdhury, Md. Mostafa Kamal Sarker, M. Monir Uddin

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a radiation oncologist, a doctor who fights cancer by painting invisible beams of energy onto a patient's body. To do this safely, you have to draw incredibly precise maps on a 3D scan of the patient's pelvis. You need to outline the tumor (so the beam hits it) and the "organs at risk" like the bladder, rectum, and small bowel (so the beam misses them).

Doing this by hand is like trying to draw a perfect map of a winding river while standing on a moving train. It takes junior doctors 2.5 hours per patient, and even experts can disagree on where the lines should go. In places like Bangladesh, where there are far more patients than doctors, this bottleneck means people wait weeks or months for treatment.

Enter BAT-RM, a new AI system that acts like a super-smart, tireless co-pilot for these doctors.

The Problem: The "Blurry Edge" and the "Huge Map"

The paper argues that previous AI attempts had two big flaws. First, they were often "blurry" at the edges. Imagine trying to trace a jagged coastline with a thick marker; you miss the tiny inlets and bays. For radiation, missing a tiny bit of the tumor or accidentally painting a healthy organ is dangerous. Second, some AI models tried to look at the entire body scan at once to understand the big picture. But looking at a whole 512x512 pixel map all at once is like trying to read a whole encyclopedia in one second—it requires so much computer power that it crashes most hospital machines.

The paper explicitly rules out the idea that standard convolutional networks (like the popular nnUNet) or full-screen Transformers are the perfect solution for this specific job. It argues that without special tricks, these models either miss the fine details or get too slow and heavy to run in a real clinic.

The Solution: A Hybrid Detective

The authors built BAT-RM, which is like a detective team with two specialists working together, plus a smart manager.

The Edge Detective (BAT Branch): This part of the AI is obsessed with boundaries. It uses a mathematical tool called a "Sobel gate" to act like a high-contrast filter. Instead of looking at the whole room, it only pays attention to the sharp edges where one organ ends and another begins. It's like using a magnifying glass specifically on the coastline to trace every tiny curve.
The Long-Range Scout (Mamba Branch): This part is the "scout" that walks through the entire map in four directions (North, South, East, West) very quickly. It uses a new technology called "Mamba" that can remember the whole journey without needing a massive memory bank. It understands that a small bowel loop on the left might be connected to one on the right, even if they are far apart.
The Smart Manager (BRAF Fusion): This is the glue. It takes the "Edge Detective's" sharp lines and the "Scout's" big-picture context and blends them perfectly. It decides, "Okay, for this part of the bladder, trust the edge; for this part of the bowel, trust the big picture."

The Proof: It Works Better and Faster

The team didn't just build this in a lab; they tested it on a massive dataset of 1,011 patients from three different hospitals in Bangladesh. They compared BAT-RM against four other top AI models (nnUNet, SegMamba, TransUNet, and UNETR).

The results were clear:

Accuracy: BAT-RM drew the maps more accurately than any of the others. For the tumor (GTV), it achieved a score of 0.966 (where 1.0 is perfect), beating the next best model.
Precision: The distance between the AI's line and the expert's line was tiny—only 1.50 mm for the tumor and 1.24 mm for the rectum. This is crucial because radiation beams are sharp; being off by even a few millimeters can hurt healthy tissue.
Speed: It runs on a standard hospital computer card (an NVIDIA T4) and takes only 18.3 milliseconds to process a single slice of the scan. A full scan takes less than 2 seconds.

The Real-World Test: The Reader Study

To see if this actually helps doctors, the authors ran a "reader study" with 13 radiation oncologists (4 experts, 4 seniors, and 5 juniors). They asked the doctors to draw maps on 100 patients twice: once alone, and once with BAT-RM's help.

The results were dramatic:

Junior doctors (who usually take 152 minutes per patient) dropped their time to just 29 minutes. That is an 80.7% reduction.
Their accuracy jumped from 0.899 to 0.965, meaning a junior doctor with AI help became almost as good as a senior doctor working alone.
The need for experts to double-check the juniors' work dropped from 52.1% to 14.3%.

The Deployment: From Paper to Patient

The paper doesn't stop at "it works in a test." The team built a web app that doctors can use right now. It connects directly to the hospital's systems (Varian, RayStation, Monaco). When a doctor approves the AI's map, it instantly exports as a standard file ready for the radiation machine.

Because of this, at the Bangladesh Medical University hospital, the time a patient waits to start treatment dropped from 1–3 days down to 2–3 hours. The paper suggests this is a real, measured change in the hospital's workflow, allowing more patients to be treated without hiring more staff.

What the Paper Says It's Not

The authors are careful to note what they haven't done yet.

They haven't proven this works for every country in the world yet; the data is from Bangladesh.
They haven't tested it on MRI scans (which are better for some tumors) yet; they only used CT scans.
They haven't done a massive, multi-country trial to prove it works everywhere, though they did test it on a separate group of 230 patients from a fourth hospital, and it still performed very well.

The Bottom Line

BAT-RM isn't just a faster computer; it's a system designed to fix the specific problems of drawing cancer maps: the blurry edges and the heavy computing load. By combining a "boundary-focused" eye with a "long-range" memory, it helps junior doctors work like experts, cutting treatment wait times from days to hours. The paper presents this as a proven, clinically deployed solution that is ready to help patients in resource-limited settings right now.

Technical Summary: BAT-RM for Cervical Cancer Radiotherapy Auto-Contouring

1. Problem Statement

Cervical cancer remains a leading cause of cancer mortality globally, with a disproportionate burden in low- and middle-income countries (LMICs) like Bangladesh, where severe shortages of radiation oncologists result in treatment wait times exceeding six months. Safe radiotherapy delivery relies on the precise manual delineation of target volumes (Gross Tumor Volume [GTV], Clinical Target Volume [CTV]) and Organs at Risk (OARs: bladder, rectum, small bowel, femoral heads). Manual contouring is time-consuming (45 minutes to 3 hours per patient) and suffers from significant inter-observer variability.

Existing deep learning solutions face three primary limitations in this context:

Boundary Precision: Convolutional architectures often lack explicit mechanisms for soft-tissue boundary delineation, leading to suboptimal precision at critical interfaces (e.g., GTV-parametrium).
Long-Range Context: Transformers address global context but incur quadratic computational complexity ( $O(n^2)$ ), making full-resolution (512×512) processing prohibitive on clinical hardware. Conversely, CNNs have limited receptive fields.
Clinical Validation Gap: Most studies rely on single-institution datasets without rigorous multi-center generalization analysis or prospective reader studies measuring workflow efficiency and patient-level impact in resource-constrained settings.

2. Methodology: BAT-RM Architecture

The authors propose BAT-RM (Boundary-Aware Transformer with Region-Aware Mamba), a hybrid 2D multi-branch architecture designed to simultaneously achieve high boundary precision and efficient long-range contextual modeling. The system processes axial CT slices (512×512) and reconstructs 3D volumes via post-processing.

Core Architectural Components

The model integrates a hierarchical CNN encoder-decoder backbone with three specialized modules:

Gated Boundary-Aware Transformer (BAT) Branch:
- Function: Explicitly models organ boundaries to resolve ambiguity at tissue interfaces.
- Mechanism: Operates on the $E_3$ encoder features (128×128 resolution). It computes a spatial gradient magnitude map using fixed Sobel operators on feature activations. This gradient map is collapsed into a learnable spatial gate ( $\mathcal{G} \in [0,1]$ ) via a 1×1 convolution and sigmoid activation.
- Attention: The gate modulates Query ( $Q$ ) and Key ( $K$ ) projections in a Strip-Pooling Multi-Head Self-Attention mechanism. This restricts attention computation to high-gradient boundary regions, reducing complexity from $O(n^2)$ to $O(H+W)$ while focusing on clinically relevant tissue interfaces (e.g., GTV-parametrium).
- Auxiliary Supervision: A boundary prediction head outputs a boundary probability map trained with Binary Cross-Entropy (BCE) loss.
Multi-Directional Recurrent Context Module (RM Branch):
- Function: Captures long-range anatomical dependencies (e.g., scattered small bowel loops) with linear computational complexity.
- Mechanism: Operates on the bottleneck features ( $E_5$ , 32×32 resolution). It employs a Mamba Selective State Space Model (SSM) with four-directional sequential scanning (North→South, South→North, East→West, West→East).
- Efficiency: By serializing the 2D feature map into four 1D sequences and processing them with Mamba's selective scanning, the module achieves $O(n)$ complexity, avoiding the memory bottlenecks of full spatial self-attention.
- Gating: SiLU-activated gates suppress uninformative tokens (e.g., homogeneous background) while amplifying anatomically relevant structures.
Boundary-Region Attention Fusion (BRAF) Module:
- Function: Adaptively combines boundary-sensitive features (from BAT) and region-aware global context (from RM).
- Mechanism: RM features are upsampled to match BAT resolution. A learnable spatial-channel attention gate ( $\alpha$ ) computes a convex combination of the two streams. The fused features are further modulated by the predicted boundary skeleton ( $\hat{B}$ ) from the BAT branch to enforce sharp edge delineation before passing to the decoder.

Training Strategy

Loss Function: A composite loss supervises segmentation overlap, boundary precision, and smoothness:
- Soft Dice Loss: Weighted by inverse voxel frequency to handle class imbalance.
- Focal Loss: Focuses training on hard, ambiguous boundary pixels.
- Boundary BCE Loss: Supervises the auxiliary boundary head.
- Smoothness Regularization: Penalizes gradient discrepancies between predicted and ground truth maps to prevent jagged contours.
Data: Trained on a multi-institutional dataset of 1,011 patients from three Bangladeshi centers (BMU, SHL, LSH), covering seven anatomical classes.

Inference Pipeline

The system employs a slice-wise 2D inference strategy followed by 3D Gaussian post-processing ( $\sigma=0.8$ mm) to ensure inter-slice continuity without the memory cost of 3D networks. The output is reconstructed into a DICOM RT-Structure Set compatible with Varian, RayStation, and Monaco treatment planning systems.

3. Key Contributions

Novel Hybrid Architecture: Introduction of BAT-RM, which uniquely combines Sobel-gated boundary attention, multi-directional Mamba scanning, and boundary-skeleton-guided fusion to address both boundary precision and long-range context simultaneously.
Large-Scale Multi-Institutional Dataset: Creation and rigorous quality control of the largest cervical cancer CT dataset from an LMIC setting (1,011 patients, three centers, seven classes), including inter-rater reliability analysis (mean Jaccard = 0.961).
Comprehensive Clinical Validation:
- Quantitative: Superior performance over four baselines (nnUNet, SegMamba, TransUNet, UNETR) across all seven classes with statistically significant improvements (Bonferroni-Holm corrected).
- Qualitative: Mechanistic evidence via error maps, contour overlays, and multi-method GradCAM analysis confirming boundary focus.
- Prospective Reader Study: A multi-center study with 13 oncologists (100 cases) demonstrating significant workflow improvements.
End-to-End Clinical Deployment: Development of a production-grade web application with native DICOM RTSTRUCT export, deployed at Bangladesh Medical University, reducing patient wait times from days to hours.

4. Results

Quantitative Performance

BAT-RM achieved state-of-the-art results on the held-out test set (254 patients):

Target Volumes: GTV Dice = 0.9662 (vs. nnUNet 0.9450); CTV Dice = 0.9571 (vs. nnUNet 0.8966).
OARs: Rectum Dice = 0.9556; Bladder Dice = 0.9655; Small Bowel Dice = 0.9741.
Boundary Accuracy: GTV HD95 = 1.500 mm; CTV HD95 = 1.507 mm; Rectum HD95 = 1.242 mm. These represent significant reductions in boundary error compared to baselines (e.g., GTV HD95 reduced by 38.7% vs. nnUNet).
Efficiency: 28.4M parameters, 41.2 GFLOPs, and 18.3 ms inference time per slice on a single NVIDIA T4 GPU (3.8 GB VRAM), significantly outperforming 3D baselines in speed and memory usage.

Reader Study Findings

Accuracy: AI assistance elevated junior oncologists' IoU from 0.899 to 0.965 (approaching senior unaided performance of 0.938).
Efficiency: Contouring time was reduced by 80.7% for juniors (from 152 to 29 minutes per case).
Workflow: Expert consultation rates for juniors dropped from 52.1% to 14.3%. Self-reported confidence scores increased by 40% for juniors.
Deployment Impact: Following deployment, patient wait times for radiotherapy simulation decreased from 1–3 days to 2–3 hours without additional staffing.

Ablation Studies

Systematic ablation confirmed the necessity of each component:

Removing the BAT block caused the largest performance drop (mIoU -0.0509), particularly for boundary-ambiguous structures like the rectum and GTV.
Removing the RM block degraded performance on structures requiring global context (bladder, femoral heads).
Removing the BRAF fusion module reduced performance across all small OARs, highlighting the importance of adaptive feature combination.

5. Significance and Claims

The paper claims that BAT-RM demonstrates a rigorous translational pipeline that moves beyond benchmark exercises to deliver measurable patient benefits in resource-constrained settings. Key claims include:

Technical Superiority: The integration of boundary-aware attention and linear-time Mamba scanning solves the dual challenges of boundary precision and long-range context without prohibitive computational costs.
Clinical Utility: The system is not merely an algorithmic improvement but a deployable tool that significantly reduces clinician workload and wait times, enabling same-day or next-day treatment initiation.
LMIC Relevance: The study provides the first multi-center reader study for cervical cancer auto-contouring in an LMIC setting, proving that AI can mitigate specialist shortages and improve care equity.
Robustness: The model generalizes across heterogeneous scanner manufacturers and protocols from three institutions and maintains performance on an independent external cohort (230 patients) without fine-tuning.

The authors conclude that this work establishes a framework for translating machine learning research into direct clinical impact, characterized by multi-center data curation, architecture design motivated by clinical failure modes, and production deployment in routine practice.