PaperlayerView exact revision on arXiv

Source-linked AI summary

Anatomy-Change-Aware Bidirectional Selective State-Space Memory for Clinically Deployed Thoracic Radiotherapy Auto-Contouring

Galib Ahmed, Istiak Ahmed, Aritra Islam Saswato, Asib Mostakim Fony, Kazi Shahriar Sanjid, Md. Tanzim Hossain, Md. Anwarul Islam, Md. Nishan Khan, Md. Misbah Khan, Labiba Faiza Karim, Jobaer Rahman, S M Hasibul Hoque, Rahnuma Shahrin Rista, Kamruzzaman Rumman, Md Arifur Rahman, Syed Md. Akram Hussain, Mohammad Ashrafuzzaman Khan, M. Monir Uddin

arXiv:2609.16036v2eess.IVcs.CV

TL;DR

Thoracic auto-contouring must handle anatomy that changes across CT slices, while manual contouring is time-intensive and operator-dependent. DAMM-Net++ uses anatomy-change-aware bidirectional memory with boundary refinement and uncertainty estimation; it reports strong segmentation and reader-study results across multicenter evaluation.

  • Problem

    Thoracic anatomy changes along the craniocaudal axis, while manual contouring is time-intensive, operator-dependent, and subject to inter-observer variability.

    Source: §1 · pp. 1–2

  • Method

    DAMM-Net++ is a 2.5D segmentation framework with bidirectional selective state-space memory conditioned on inter-slice anatomical change, plus boundary refinement and uncertainty estimation.

    Source: §1 · p. 2

  • Results

    Mean DSC was 0.955 and HD95 was 3.78 mm; AI assistance reduced contouring time by 75–80% and raised junior-reader IoU from 0.861 to 0.925.

    Source: §1 · p. 2, §8 · p. 13

  • Takeaways & Limitations

    The uncertainty head is well-calibrated and supports case-level triage, concentrating residual error where expert review is most needed.

    Source: §8 · p. 13

    The dataset is geographically concentrated in Bangladesh, and inference across experience groups relies on 17 independent readers.

    Source: §7 · pp. 12–13

Abstract

from arXiv · show

We developed DAMM-Net++, a 2.5D architecture for thoracic OAR and target volume segmentation that addresses three persistent challenges in radiotherapy auto-contouring: inter-slice surface incoherence, systematic failure on small low-contrast targets, and the absence of per-case reliability signals. The central component is an anatomy-change-aware bidirectional selective state-space memory that models through-plane anatomical change and selectively propagates context along the axial slice sequence. A boundary-aware decoder sharpens near-surface predictions, and an uncertainty head provides calibrated per-voxel confidence for clinical triage. We evaluated 2,146 patients across four centers, an independent external cohort of 112 patients, and a multicenter reader study involving 17 radiation oncologists on 305 cases. The model achieves a mean Dice of 0.955 and HD95 of 3.78 mm, with the largest gains on low-contrast organs-at-risk (OARs) and target volumes where through-plane context is most critical. The uncertainty head is well-calibrated and supports case-level triage. In the reader study, AI assistance reduced contouring time by 75-80% across experience levels and raised junior-reader IoU from 0.861 to 0.925, matching the unedited model. External validation showed a modest internal-to-external drop (< 5%) with calibrated uncertainty transferring without recalibration. The complete deployment pipeline from DICOM ingestion to TPS-compatible RTSTRUCT export has been integrated into the clinical workflow at a partner hospital, where it is used to assist with contouring. These results suggest that anatomically motivated inter-slice memory, paired with uncertainty-guided review, offers a clinically viable path for thoracic auto-contouring.

1. Introduction

Thoracic radiotherapy contouring is challenged by changing anatomy across slices, limitations in existing 2.5D memory, and ambiguous low-contrast targets. DAMM-Net++ combines anatomy-change-aware bidirectional memory with boundary refinement and uncertainty estimation, and is evaluated through multicenter benchmarking and reader studies.

1. Introduction

  • Conventional ConvLSTM and ConvGRU transitions are fixed and do not adapt to anatomical changes between individual slices.The motivation is especially relevant to structures with rapid or irregular appearance changes and to low-contrast GTV and CTV segmentation.Source: §1 · p. 2
  • Evaluation includes controlled real-versus-randomized transition experiments, external validation, uncertainty analysis, reader evaluation, and clinical workflow integration.The benchmark spans 2,146 patients from four centers and includes external validation on 112 patients; the pipeline exports TPS-compatible RTSTRUCT files.Source: §1 · p. 2
  • DAMM-Net++ introduces anatomy-change-aware bidirectional selective state-space memory in a 2.5D thoracic CT framework.The model conditions state-space dynamics on observed inter-slice transitions and processes context in both axial directions.Source: §1 · p. 2
  • A memory-guided boundary-aware decoder and single-pass uncertainty head provide boundary refinement and case-level predictive quality assessment.Source: §1 · p. 2
  • A multicenter reader evaluation with 17 radiation oncologists demonstrated 75–80% reductions in contouring time and increased junior-reader IoU from 0.861 to 0.925.Source: §1 · p. 2

2. Related work

Prior work addresses auto-contouring, through-plane context, selective state-space models, and auxiliary components, but the supplied review identifies limits in anatomical adaptation and clinical quality assurance.

2.1. Thoracic radiotherapy auto-contouring

  • Auto-contouring can improve inter-observer consistency, but target boundaries and low-contrast esophagus remain difficult; the largest gains are reported for trachea and esophagus.A prospective five-center study is also cited as reporting improved geometric agreement and reduced contouring time with AI-assisted OAR delineation.Source: §2.1 · pp. 2–3

2.2. Through-plane context modeling

  • 2D methods discard inter-slice information, whereas volumetric 3D methods integrate context at substantial memory cost.2.5D recurrent methods process adjacent slices, but fixed transition operators cannot adapt to each observed transition and sequential updates complicate full-thorax training.Source: §2.2 · p. 3

2.3. Selective state-space models

  • DAMM-Net++ conditions selective state-space dynamics on a multi-scale representation of actual inter-slice anatomical change and runs the memory bidirectionally.The review states that prior medical imaging studies had not explicitly conditioned selective state-space dynamics on this representation in this setting.Source: §2.3 · p. 3

2.4. Auxiliary components

  • Boundary-aware supervision emphasizes near-surface voxels relevant to dose calculations, while uncertainty calibration and case-level triage require explicit evaluation.Single-pass aleatoric uncertainty is presented as an efficient alternative to expensive Bayesian approximations.Source: §2.4 · p. 3

3. Methods

The study uses a retrospective, multicenter thoracic planning CT dataset with clinically approved delineations of nine structures, then prepares axial-slice inputs for training and evaluation.

3.1. Dataset preparation

  • 2,146 patients from four Bangladeshi tertiary hospitals over approximately eight years formed the retrospective dataset.Cases with truncation, severe artifacts, or incomplete contours were excluded; CT volumes had mean inter-slice spacing of 2.493 ± 0.097 mm.Source: §3.1 · p. 3
  • Board-certified radiation oncologists produced RTSTRUCT contours under ESTRO and RTOG guidelines, with harmonized nomenclature and geometric validation.An inter-rater study of 100 cases with ten oncologists found IoU exceeding 0.95 for every structure.Source: §3.1 · p. 3
  • The internal cohort of 2,034 patients was institution-stratified into training, validation, and test sets, while 112 UHL patients were reserved for external validation.The internal split was 70/15/15; Table 1 identifies BMU, SHL, and LH as internal institutions and UHL as external.Source: §3.1 · p. 3, Table 1 · p. 4
  • CT preprocessing mapped intensities to [0, 255], resampled images to 512 × 512, and decomposed volumes into axial slices for training and evaluation.Only axial slices were used, matching the standard clinical contouring view.Source: §3.1 · p. 4

3.2. Model architecture

DAMM-Net++ processes consecutive axial CT slices through a shared encoder, anatomy-change-aware bidirectional memory, and a memory-guided boundary-aware decoder. Its task heads produce segmentation and related boundary, distance, presence, and confidence outputs, trained with a multi-task objective.

3.2. Model architecture

  • DAMM-Net++ takes five consecutive axial slices through a shared ConvNeXt encoder, inter-slice transition branch, bidirectional state-space memory, and memory-guided boundary-aware decoder.The architecture produces segmentation, boundary, signed-distance, presence, and uncertainty outputs.Source: §3.2 · p. 4, Figure 1 · p. 5

3.2.1. ConvNeXt encoder

  • The ConvNeXt encoder processes slices independently into a four-level feature pyramid, reserving through-plane reasoning for downstream transition and memory modules.Its feature maps have resolutions H/4, H/8, H/16, and H/32; the three shallower maps provide decoder skip connections.Source: §3.2.1 · p. 4

3.2.2. Multi-scale inter-slice transition branch

  • The transition branch computes adjacent-slice bottleneck feature differences, then processes them at three scales with channel and spatial attention.A learnable, zero-initialized residual combines the refined transition with the original difference, covering fine- and coarse-scale anatomical changes.Source: §3.2.2 · p. 4

3.2.3. Bidirectional selective state-space memory

  • The memory module adds the anatomy-change-enhanced transition to bottleneck features before input-dependent state-space dynamics, rather than simply concatenating the transition signal.Independent forward and backward sweeps are fused with a spatial gate; cross-attention then uses fused memory to enhance current bottleneck features.Source: §3.2.3 · p. 5

3.2.4. Memory-guided boundary-aware decoder

  • Memory-guided skip attention filters encoder skip features, while boundary-aware refinement amplifies features near organ edges during decoder upsampling.Intermediate boundary predictions from three decoder levels are fused at full resolution.Source: §3.2.4 · p. 6

3.2.5. Task heads and uncertainty estimation

  • Task-specific heads predict segmentation, signed distance, organ presence, and per-pixel aleatoric uncertainty, from which a confidence map is derived.The deployment interface flags pixels with confidence below 0.7 for manual review.Source: §3.2.5 · p. 6

3.2.6. Multi-task loss

  • The training objective combines eight weighted loss terms, including Dice and focal losses with uncertainty reduction, boundary and deep-supervision losses, and spatial-continuity and presence terms.A regularization term discourages the trivial all-high-variance solution.Source: §3.2.6 · p. 6

3.3. Experimental setup

The experiments compare DAMM-Net++ with four baselines under matched internal data splits and preprocessing, while retaining each model’s intended configuration. Evaluation uses reconstructed 3D label volumes, complementary structure-level averages, and paired, multiplicity-adjusted statistical tests.

3.3. Experimental setup

  • DAMM-Net++ was compared with nnUNet, nnMamba, UNETR++, and SwinUNETR using identical internal splits and preprocessing, while each baseline retained its native configuration.DAMM-Net++ used its intended 2.5D configuration; the other models used their published volumetric configurations.Source: §3.3 · p. 6
  • DAMM-Net++ training used five-slice sequences, batch size two sequences, 512 × 512 inputs, and 200 epochs.Training used AdamW with a 10−4 initial learning rate and 10−5 weight decay, plus cosine annealing warm restarts and standard augmentations.Source: §3.3 · pp. 6–7
  • All models were evaluated on reconstructed 3D label volumes in each patient’s native voxel grid, avoiding a two- versus three-dimensional comparison confounder.DAMM-Net++ slice predictions were assembled into probability volumes, decoded, and mapped to native coordinates.Source: §3.3 · p. 7
  • Results include per-structure mean ± SD and 95% confidence intervals, plus macro- and voxel-weighted micro-averages across the nine structures.The macro-average gives each structure equal weight, including small, low-contrast structures.Source: §3.3 · p. 7
  • Per-patient comparisons used paired Wilcoxon signed-rank tests with Holm–Bonferroni adjustment, and effect sizes were reported as Hedges’ g with BCa bootstrap intervals.The adjusted two-tailed significance level was α = 0.05.Source: §3.3 · p. 7

4. Ablation study

The ablations attribute DAMM-Net++’s largest component gains to through-plane memory and inter-slice transitions, with benefits concentrated on anatomically variable structures. Its uncertainty head chiefly improves calibration and supports selective review rather than substantially increasing overlap.

4.1. Component contribution

  • The selective SSM memory and multi-scale inter-slice transition branch yield the two largest component gains; the uncertainty head adds only marginal overlap improvement.Table 2 adds components sequentially; the uncertainty head’s value is calibration rather than Dice.Source: §4.1 · p. 7, Table 2 · p. 8

4.2. Controlled ablation of the inter-slice transition signal

  • Real inter-slice transitions significantly outperform randomized and shuffled controls (p<0.001, Hedges’ g>0.8), supporting the value of anatomical-change information.The randomized control performs no better than SSM-only; shuffled transitions show a modest, non-significant improvement.Source: §4.2 · p. 7, Table 3 · p. 8

4.3. Anatomical locus of the gains

  • Gains over the static baseline are negligible for body and lungs but several Dice points for the variable, low-contrast esophagus, trachea, GTV, and CTV.The comparison contrasts Base without transition and memory modules with the full configuration.Source: §4.3 · p. 7, Table 4 · p. 8

4.4. Uncertainty estimation and calibration

  • The uncertainty-attenuated loss improves calibration and error detection, while flagging 5% of voxels raises retained-region Dice above 0.98.Calibration and error detection outperform a plain variance head or softmax-entropy proxy; low-confidence voxels are routed to manual review.Source: §4.4 · p. 7, Table 5 · p. 8
  • Selective SSM outperforms recurrent, attention-based, and non-selective SSM alternatives on accuracy–latency trade-off; chosen defaults are state dimension 16, sequence length 5, and three transition scales.Multi-scale processing, channel-spatial attention, memory-guided skip attention, and boundary-aware refinement each contribute positively, with the largest gains on NSD and HD95.Source: §4.4 · p. 8

5. Results

DAMM-Net++ leads the reported multi-structure comparison in overlap and distance metrics, with especially strong results on target volumes. Its uncertainty estimates support case-level review, while qualitative analyses show coherent volumes and boundary-localized errors.

5.1.1. Overlap and boundary performance

  • CTV and GTV DSC are 0.936 ± 0.019 and 0.957 ± 0.012, respectively, versus baseline ranges of 0.924–0.928 and 0.942–0.946; CTV HD95 is 2.80±1.83 mm versus baseline values above 10.6 mm.The Heart illustrates that similar volumetric overlap can differ in boundary placement, while Body performance is near-ceiling across models.Source: §5.1.1 · pp. 9–10

5.1.2. Uncertainty quantification and case-level

  • Across 2,745 predictions, weighted mean absolute calibration error is 0.0095; a three-feature classifier predicts IoU < 0.85 with pooled AUROC 0.960.Sparsification gains are largest for CTV, esophagus, trachea, and GTV, supporting review of target volumes and esophagus.Source: §5.1.2 · p. 10

5.1.3. Computational analysis and accuracy–cost trade-off

  • At 22.4 ms per axial slice on an NVIDIA T4, DAMM-Net++ is about 3.2–3.4× faster than UNETR++ and SwinUNETR, respectively, while improving DSC by +0.011 over nnUNet.The comparison reports models in native modes, with DAMM-Net++ in 2.5D and baselines in 3D.Source: §5.1.3 · p. 10

5.2.1. Spatial prediction and error analysis

  • Predictions remain spatially coherent across axial, coronal, and sagittal views; DAMM-Net++ errors form a compact boundary fringe, unlike baseline interior or irregular errors.Representative GTV overlays show closer agreement at the inferior and lateral margins, and error maps identify interior gaps in GTV or CTV predictions as under-coverage risk.Source: Figure 4 · p. 9, §5.2.1 · pp. 10–11, Figure 6 · p. 11

5.1. Quantitative analysis

  • DAMM-Net++ has the highest macro-average DSC (0.9545) and IoU (0.9144), and macro-average HD95 of 3.778 mm versus 5.65–6.11 mm for baselines.It has a significant advantage over all baselines on at least one overlap and one distance metric for eight of nine structures; Heart is the overlap exception and trachea HD95 the distance exception.Source: §5.1 · p. 8, Figure 3 · p. 9, Table 6 · p. 11

6. Clinical Evaluation, External Validation, and Deployment

AI assistance improved contour quality and reduced contouring time across reader experience levels, while external validation, uncertainty analysis, and clinical deployment assessed broader readiness.

6. Clinical Evaluation, External Validation, and Deployment

  • Seventeen oncologists across three centers evaluated 305 held-out patients under No-AI and AI-Assisted conditions, with a three-oncologist panel establishing reference contours.The study used a ≥4-week washout and designated contouring time as its primary workflow outcome.Source: §6 · pp. 11–12, Table 7 · p. 12
  • 75–80%: AI assistance reduced contouring time across reader groups and improved junior mean IoU from 0.8606 to 0.9245.IoU also increased for seniors and experts; all 17 readers benefited.Source: §6 · p. 12
  • Mean IoU decreased from 0.9322 internally to 0.9076 externally, a relative gap of −2.64%, with no structure decreasing more than 5%.The external cohort comprised 112 patients withheld from model development.Source: §6 · p. 12
  • Weighted absolute calibration error was 0.020, and uncertainty remained well calibrated on external validation.The calibration result is reported with Supplementary Fig. 5c,d and Supplementary Table 7.Source: §6 · p. 12
  • The four-stage DICOM-to-RTSTRUCT pipeline processed a median 196-slice volume in under 20 seconds on an NVIDIA T4 16-GB GPU.The clinician-in-the-loop system is deployed at Bangladesh Medical University, where contours receive radiation-oncologist review and approval before use.Source: §6 · p. 12

7. Limitations and future directions

The authors identify geographic concentration, target-volume variability, and dependence of workflow gains on local clinical conditions as boundaries on generalization and use.

7. Limitations and future directions

  • The dataset remains geographically concentrated in Bangladesh, so validation across broader populations, imaging platforms, and healthcare systems is needed.The authors also propose prospective evaluation across additional centers in South and Southeast Asia.Source: §7 · pp. 12–13
  • Target-volume delineation remains subject to inter-observer variability, requiring continued expert oversight in radiotherapy planning.Source: §7 · p. 13
  • Observed workflow gains may not translate directly to other clinical environments because staffing patterns, caseloads, and treatment-planning infrastructure vary.Source: §7 · p. 13
  • Reader-level inference rests on 17 independent readers, including five experts and five seniors, despite 10,370 case-condition observations.Future work includes PET and MRI extensions, adaptive per-structure uncertainty thresholds, and downstream dose-volume histogram evaluation.Source: §7 · p. 13

8. Conclusion

DAMM-Net++ combines anatomy-change-aware inter-slice memory with uncertainty estimates for thoracic auto-contouring, and the paper reports strong segmentation, reader-study, and external-validation results alongside clinical workflow integration.

8. Conclusion

  • 0.955 mean DSC and 3.78 mm HD95 were achieved on 2,146 multi-institutional patients.Source: §8 · p. 13
  • The largest gains were on CTV, GTV, esophagus, and trachea, where through-plane context is most critical.Source: §8 · p. 13
  • 75–80% time reduction across experience levels accompanied junior-reader IoU rising from 0.861 to 0.925, matching the unedited model.Source: §8 · p. 13
  • External validation on 112 patients showed drops below 5%, with calibrated uncertainty transferring without recalibration.Source: §8 · p. 13
  • The complete DICOM-ingestion-to-RTSTRUCT pipeline is integrated into a partner hospital workflow for contouring assistance, while geographic concentration and subjective target delineation remain limitations.The authors propose multimodality extensions and adaptive uncertainty thresholds for future work.Source: §8 · p. 13

Data and Code Availability

The dataset is restricted but may be requested under conditions, while the source code and demo are publicly accessible.

Data and Code Availability

Supplementary Material

The supplementary material details statistical and efficiency analyses, uncertainty and reader-study evaluations, ablations, external validation, and the DICOM-to-NIfTI preparation pipeline.

S1. Extended Statistical Analysis: Full Pairwise Comparisons

  • Supplementary comparisons cover paired statistical tests and effect sizes across baselines, plus computational efficiency measurements on an NVIDIA T4 GPU.The efficiency table specifies that lower values are better; the supplied caption does not give the measured values.Source: Table 1 · p. 16, Table 2 · p. 16

S3. Uncertainty Calibration and Sparsification Analysis

  • The uncertainty figure presents confidence calibration and sparsification curves against oracle and random baselines, while the multi-planar example shows coherent contours across axial, coronal, and sagittal views.The figure captions describe analyses and an illustrative case, without reporting numerical calibration or sparsification outcomes.Source: Figure 1 · p. 17, Figure 2 · p. 18

S5. Additional Ablation Studies

  • Ablations identify selective SSM as the best accuracy-latency trade-off, while decoder components contribute positively, especially on surface metrics.The tested defaults (d_state = 16, T = 5, three transition scales) achieved the highest Dice across the hyperparameter settings.Source: Table 3 · p. 19, Table 4 · p. 19, Table 5 · p. 19

S6. Consolidated Reader Study Analysis

  • The reader analysis reports improvements in IoU and HD95 for all 17 readers with AI assistance, plus lower junior consultation and expert no-edit rates.The supplementary analysis also presents per-tier effect sizes and per-class comparisons against the unedited model.Source: Figure 3 · p. 20, Figure 4 · p. 21, Table 6 · p. 21

S7. Per-Class Reader-Tier Comparison

  • Per-class reader comparisons show DAMM-Net++ reaches junior-tier IoU on all structures and yields its largest junior HD95 improvements on CTV, GTV, and trachea.Junior variability is widest, particularly for low-contrast targets.Source: Figure 4 · p. 21

S8. External Validation Analysis

  • External mean IoU decreased from 0.932 internally to 0.908 on the UH cohort, with no structure exceeding a 5% drop; external weighted absolute calibration error was 0.020.HD95 increased modestly for all structures, with the largest absolute increase on GTV (+0.77 mm).Source: Figure 5 · p. 22, Table 7 · p. 22

S9. Volumetric 3D Reconstruction

  • The supplementary material includes a volumetric reconstruction of nine structures from sixteen viewing angles and a three-panel orthogonal review application with visibility and opacity controls.It also lists the training hyperparameters used for DAMM-Net++.Source: Figure 6 · p. 23, Figure 7 · p. 23, Table 8 · p. 24

S12. DICOM-to-NIfTI Conversion and Multi-Planar Slice Extraction

  • The preparation pipeline converts DICOM CT and RTSTRUCT inputs into CT and multi-class segmentation NIfTI volumes, then saves axial, coronal, and sagittal image and label slices.It loads data in 3D Slicer, skips empty structures, assigns canonical labels, and applies CT windowing and normalization.Source: S12. DICOM-to-NIfTI Conversion and… · p. 24
Loading 2609.16036v2…