Paper deep dive
Preserving Anatomical Continuity: Three-Stage Pipeline for Colon Segmentation in 3D Abdominal CT Scans
Deshan Kalupahana, Sonit Singh, Praveen Ravindran, Arcot Sowmya
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Accurate colon segmentation from CT images is essential for colorectal disease analysis, yet deep learning based methods often produce disconnected predictions due to complex anatomy. This study introduces a three-stage, topology-preserving segmentation pipeline to address this issue. The first stage performs initial deep learning-based segmentation, followed by centreline bridging to reconnect disjoint regions and a reconstruction stage to refine continuity. Evaluations on TotalSegmentator and RAOS datasets using overlap, distance and topology-based metrics demonstrate improved structural consistency while maintaining segmentation accuracy. The proposed method enhances topological integrity, enabling more reliable colon segmentation for clinical and research applications.
Tags
Links
- Source: https://arxiv.org/abs/2610.03467v1
- Canonical: https://arxiv.org/abs/2610.03467v1
Trouble viewing inline? Open PDF directly →
Full Text
23,051 characters extracted from source content.
Expand or collapse full text
Preserving Anatomical Continuity: Three-Stage Pipeline for Colon Segmentation in 3D Abdominal CT Scans Deshan Kalupahana⋆ Sonit Singh⋆ Praveen Ravindran† Arcot Sowmya⋆ Abstract Accurate colon segmentation from CT images is essential for colorectal disease analysis, yet deep learning based methods often produce disconnected predictions due to complex anatomy. This study introduces a three-stage, topology-preserving segmentation pipeline to address this issue. The first stage performs initial deep learning-based segmentation, followed by centreline bridging to reconnect disjoint regions and a reconstruction stage to refine continuity. Evaluations on TotalSegmentator and RAOS datasets using overlap, distance and topology-based metrics demonstrate improved structural consistency while maintaining segmentation accuracy. The proposed method enhances topological integrity, enabling more reliable colon segmentation for clinical and research applications. Index Terms: Colorectal Cancer, Colon Segmentation, CT Imaging, Deep Learning, Diverticular Disease, Topology Preservation †address: ⋆School of Computer Science and Engineering, University of New South Wales, Sydney, Australia †Sydney Adventist Hospital, Wahroonga, Australia 1 Introduction Structural abnormalities of the colon are key indicators of serious gastrointestinal disorders such as colorectal cancer and diverticular disease. Colorectal cancer is among the most prevalent malignancies worldwide, with approximately 1.9 million new cases and over 900,000 deaths reported in 2022 [15]. Diverticular disease also represents a major colorectal health concern, affecting up to 70% of adults over 80 years in Western populations [11]. These conditions highlight the need for early and accurate identification of colon abnormalities. Computed tomography (CT) imaging remains the primary modality for detecting such diseases, making colon segmentation a crucial preliminary step in in-depth analysis of the region of interest for accurate identification of these conditions. With advancements in deep learning (DL)-based segmentation, automated colon analysis has become increasingly feasible. Early colon segmentation methods relied on thresholding techniques applied to cleansed bowel scans with contrast-enhancing fluids [2]. These approaches were later enhanced through DL models capable of addressing partial volume effect [4]. More recently, several studies have focussed on DL-only segmentation of the colon and sigmoid regions [12, 14, 9]. Wang et al. introduced BowelNet, a two-stage network designed for both colon and small intestine segmentation, achieving an average Dice score of approximately 81% on a private dataset for colon segmentation [12]. The TotalSegmentator framework, trained on 104 anatomical structures including the colon, recorded an 89.6% accuracy in colon segmentation, yet it also revealed ongoing challenges with discontinuous predictions [13]. This study aims to resolve the problem of discontinuous predictions by introducing a three-stage, topology-preserving colon segmentation pipeline that ensures structural continuity across disconnected areas while maintaining segmentation accuracy. 2 Method Figure 1: Three-stage pipeline for topology-preserving colon segmentation: (1) initial segmentation with possible disconnections (2) centreline bridging and (3) disconnected region prediction and merging. 2.1 Overview of the Segmentation Pipeline The proposed topology-preserving colon segmentation pipeline, illustrated in Fig. 1, comprises three stages. In Stage 1, the initial colon segmentation is performed, which may contain disconnected regions. Stage 2 constructs a centreline bridge where necessary, based on the extracted centreline from the initial segmentation. Finally, Stage 3 predicts the disconnected regions using the centreline bridge and merges them with the initial segmentation to produce a complete and topologically consistent colon structure. 2.2 Stage 1 : Initial Colon Segmentation This stage focuses on initial colon segmentation using various segmentation networks. Several UNet-based architectures were assessed, including high-performing medical image segmentation model nnUNet [3], the state-of-the-art bowel segmentation model BowelNet [12], conventional UNet variants (VNet, UNet++) [3], transformer-based models (SwinUNeTR, UNeTR, CoTR) [3], and the state-space-based variant U-Mamba [6]. All models were trained to minimise a combination of cross-entropy (CE) loss and Dice loss between the predicted segmentation and the ground truth (GT) annotations. 2.3 Stage 2 : Centreline Bridging The centreline bridging process, illustrated in Fig. 1, begins by processing the outputs from Stage 1 to identify connected and disconnected segments. First, a skeleton is generated from the predicted segmentation (Fig. 1 Stage 2 (a)). This skeleton is then converted into a graph representation, where the longest path in each connected component is defined as the main branch, and shorter branches are attached accordingly. To obtain a smooth and topologically consistent centreline, shorter branches, primarily formed due to haustral folds in the colon, are removed. This refinement reduces noise and narrows the potential endpoints for reconnection. Endpoints are identified as nodes with a single degree in the graph. Colon-specific anatomical knowledge is then applied to differentiate between the cecum and rectum endpoints, which are typically located at the bottom-left and bottom regions of the volume respectively (Fig. 1 Stage 2 (b)). Starting from the component containing the rectum endpoint, the algorithm iteratively searches for the nearest endpoint in disconnected components, progressing toward the cecum endpoint to complete the bridging (Fig. 1 Stage 2 (c)). This stage does not involve any learning-based operations, instead it follows a topology-driven approach. Connections between the identified endpoints are created using Bézier curves, constructed according to the directional vectors of the corresponding endpoints (Fig. 1 Stage 2 (c)). For the next stage, only the centreline bridges are forwarded rather than the entire connected centreline. These bridges include extended points derived from the original centreline, enabling the generation of disconnected regions that remain in contact with the initial segmentation (Fig. 1 Stage 2 (d)). 2.4 Stage 3 : Reconstruction of Full Colon In this stage, a UNet model trained using the ground truth centreline and corresponding CT volume, optimised with CE and Dice losses, was employed to predict the disconnected regions. Since only the centreline bridges are provided as input, the network focusses on reconstructing those specific missing segments. To mitigate potential overpredictions, the predicted components are filtered based on their connectivity to the designated endpoints. Finally, the filtered disconnected regions are merged with the initial segmentation to produce a complete and continuous colon reconstruction. 2.5 Experimental Setup Two CT multi-organ datasets containing colon structures were used in this study. The first dataset is TotalSegmentator (TS) and originally comprised 1,024 samples, from which 467 abdominal scans containing the colon were selected [13]. The second dataset, RAOS, included 413 abdominal CT samples. TS was annotated using an active learning approach, while RAOS was manually annotated [5]. In Stage 1, the best-performing UNet variant for colon segmentation was identified and subsequently used to evaluate the complete three-stage pipeline. For both datasets, the models were trained using five-fold cross-validation with identical data splits to ensure consistency. The outputs from Stage 1 were manually reviewed and grouped to provide a clearer assessment of the pipeline’s performance, enabling analysis of qualitative improvements while ensuring that the initial segmentation accuracy remained unaffected. Segmentation performance was evaluated using overlap, distance, and topology-based metrics. Overlap metrics included Dice and IoU [8], assessing volumetric agreement with the GT annotation. Distance metrics comprised Hausdorff Distance [8] and Structural Similarity Index (SSIM) [1], measuring boundary deviation and structural similarity respectively. Topology-based metrics included Centreline Dice (clDice) [10], Average Centreline Distance (ACD) [7] and Connectivity Ratio (CR), which evaluate continuity, geometric alignment, and structural consistency of the predicted colon segmentation. The connectivity ratio is defined as |1−B0predB0GT| |1- B_0^predB_0^GT |, where B0predB_0^pred and B0GTB_0^GT denote the Betti-0 numbers representing the number of connected components in the prediction and ground truth, respectively. Overlap-based metrics are often insensitive to small disconnections that minimally affect region overlap, but can disrupt anatomical correctness. In contrast, topology-based metrics are crucial for evaluating the continuity of the colon and ensuring structurally coherent segmentations. 3 Results and Analysis 3.1 Initial Colon Segmentation Table 1: Dice scores of Stage 1: Initial Colon Segmentation UNet Variant TotalSegmentator RAOS nnUNet 0.9308±0.0631 0.8847±0.0706 BowelNet 0.9041±0.0848 0.8457±0.0929 VNet 0.8882±0.0889 0.8348±0.0797 UNet++ 0.9118±0.0722 0.8594±0.0735 SwinUNeTR 0.8701±0.0931 0.8513±0.0741 UNetTR 0.8678±0.0892 0.8200±0.0803 CoTr 0.9103±0.0802 0.8634±0.0886 U-mamba 0.9185±0.0740 0.8788±0.0738 In Table 1, the results of Stage 1 for both datasets are presented. The nnUNet achieved the best performance for colon segmentation across both datasets, while the other variants performed comparably, with only marginal differences. Qualitative analysis revealed that nearly half of the predicted segmentations closely aligned the ground truth annotations (TS: 49.25%, RAOS: 52.78%). However, a subset of the predictions contained disconnected regions (TS: 18.41%, RAOS: 23.48%), while in some cases, the predicted results were visually more accurate than the ground truth (TS: 9.20%, RAOS: 18.40%). Both situations contributed to a reduction in the Dice score. Disconnections typically occurred due to the loss of fine structural details in the CT volumes. Additionally, the patch-based training approach and the use of Dice loss, which optimises for regional overlap rather than structural continuity, may have limited the network’s ability to preserve the overall colon topology. Table 2: Experiment results of TS dataset Data Category All samples Samples where disconnections visible after Stage 1 All samples with disconnections after Stage 1 Samples having continues GT, but disconnections after Stage 1 All <120m <90m No. of Samples 467 86 46 33 27 Metrics Stage 1 Stage 3 Stage 1 Stage 3 Stage 1 Stage 3 Stage 1 Stage 3 Stage 1 Stage 3 ↑ Dice 0.9308±0.0631 0.9285±0.0672 0.8708±0.1077 0.8619±0.1137 0.8922±0.0756 0.8885±0.0763 0.9145±0.0453 0.9126±0.0480 0.9202±0.0471 0.9201±0.0494 ↑ IoU 0.8759±0.0932 0.8726±0.0978 0.7847±0.1443 0.7720±0.1492 0.8128±0.1111 0.8068±0.1121 0.8455±0.0745 0.8426±0.0786 0.8554±0.0771 0.8556±0.0803 ↑ ClDice 0.9551±0.0686 0.9517±0.0740 0.8849±0.1100 0.8725±0.1185 0.9028±0.0826 0.9010±0.0893 0.9344±0.0498 0.9362±0.0546 0.9456±0.0459 0.9509±0.0461 ↑ SSIM 0.9947±0.0039 0.9946±0.0041 0.9929±0.0044 0.9923±0.0048 0.9937±0.0038 0.9934±0.0042 0.9947±0.0028 0.9946±0.0029 0.9948±0.0026 0.9948±0.0026 ↓ HD95 5.3258±9.3093 5.5621±9.7934 12.8904±15.3848 13.2122±15.8440 10.0087±8.8628 8.8088±7.9481 6.6820±6.1570 6.0453±5.5801 5.5017±5.7629 5.0630±5.5834 ↓ ACD 1.9403±1.8081 1.9940±1.8883 3.5816±2.8222 3.7196±2.9436 3.0152±1.9857 2.7998±1.8404 2.2358±1.1202 2.1148±1.0324 2.0948±1.1764 1.9650±1.0523 ↓ CR 0.0974±0.1387 0.0918±0.1536 0.1974±0.2553 0.1637±0.3029 0.1867±0.1460 0.1260±0.1149 0.1343±0.1074 0.1090±0.0953 0.1254±0.1127 0.1126±0.1033 Table 3: Experiment results of RAOS dataset Data Category All samples Samples where disconnections visible after Stage 1 All samples with disconnections after Stage 1 Samples having continues GT, but disconnections after Stage 1 All <120m <90m No. of Samples 413 97 47 41 36 Metrics Stage 1 Stage 3 Stage 1 Stage 3 Stage 1 Stage 3 Stage 1 Stage 3 Stage 1 Stage 3 ↑ Dice 0.8847±0.0706 0.8825±0.0717 0.8519±0.0885 0.8443±0.0885 0.8634±0.0789 0.8589±0.0781 0.8642±0.0808 0.8625±0.0790 0.8685±0.0717 0.8689±0.0693 ↑ IoU 0.7994±0.1001 0.7962±0.1015 0.7508±0.1165 0.7393±0.1155 0.7671±0.1107 0.7599±0.1092 0.7687±0.1132 0.7658±0.1106 0.7739±0.1040 0.7741±0.1009 ↑ ClDice 0.9078±0.1031 0.9017±0.1080 0.8602±0.1241 0.8389±0.1270 0.8754±0.1123 0.8566±0.1157 0.8694±0.1162 0.8588±0.1216 0.8704±0.1149 0.8650±0.1202 ↑ SSIM 0.9876±0.0056 0.9873±0.0058 0.9864±0.0059 0.9856±0.0063 0.9861±0.0057 0.9854±0.0058 0.9859±0.0061 0.9854±0.0062 0.9857±0.0062 0.9853±0.0063 ↓ HD95 10.4187±14.5159 10.6874±14.7654 16.2218±18.8779 17.0241±19.4622 12.0387±12.4027 12.1059±11.9316 11.6243±11.6203 11.2389±10.9607 10.9384±10.0858 10.3057±9.6813 ↓ ACD 4.5196±2.6928 4.5763±2.6889 5.5578±3.8153 5.7576±3.7254 4.8508±2.5429 4.9573±2.3608 4.8731±2.6435 4.8631±2.4374 4.7193±2.2952 4.7378±2.3083 ↓ CR 0.3234±0.1391 0.3082±0.1400 0.3464±0.1650 0.2919±0.1668 0.3653±0.1668 0.3143±0.1695 0.3540±0.1736 0.3131±0.1752 0.3447±0.1718 0.3092±0.1785 3.2 Final Colon Reconstruction In Tables 2 and 3, the experimental results obtained after Stage 3 are presented. The evaluation was performed by grouping the samples according to the quality of the Stage 1 predictions. The first category includes all samples (columns 2 and 3). The slight decrease in quantitative evaluation metrics is mainly attributed to cases where the predictions either overpredicted or improved upon the GT annotations. These improvements occur when the pipeline successfully reconstructs missing regions in GT annotations that contain disconnections. The second category includes samples that exhibited disconnections in Stage 1 predictions, excluding the best overlapping cases (columns 4 and 5). This category shows a noticeable drop in overlap metric values compared to Category 1. As some of these samples also contain disconnections in the GT, the quantitative improvement is not clearly reflected in the final segmentation results. Specifically, the TS dataset contained 40 ground truth annotations with disconnections, while the RAOS dataset contained 50 such cases. The third category considers samples where the GT annotations were continuous, but the Stage 1 predictions showed disconnections (columns 6 and 7). This category allows us to quantitatively assess the improvements introduced by the proposed method. The results show notable gains in topology-based metrics, while slight decreases are observed in overlap and distance-based evaluations. These decreases primarily occur in cases with large disconnections where the predicted reconnections do not correspond precisely to the GT annotations. To address this, a distance threshold was introduced, connecting endpoints only when the Euclidean distance between them was less than 120m (columns 8 and 9) and, in a refined setting, less than 90m (columns 10 and 11), based on an analysis of all samples. Under these thresholds, improvements are observed across most metrics, especially for disconnections below 90m. However, the quantitative improvements remain small, as the reconnected regions are relatively small and therefore contribute only minimally to the overall metric gains. 3.3 Qualitative analysis Figure 2: Samples from the TotalSegmentator and RAOS datasets showing cases with disconnections, and the progressive reconstruction of colon segmentation across pipeline stages compared to the ground truth (GT) annotations. Visual inspection of the results provides important insights into the effectiveness of the proposed method, clearly highlighting the improved continuity and reconnection of colon segments that are not fully reflected by the quantitative metrics. In Fig. 2 sample results from the TS and RAOS datasets after applying the proposed pipeline are shown. The reconnection process performs effectively within the defined disconnection distance threshold for both datasets. However, a few failure cases were observed in the centreline bridging stage. In rare instances, the algorithm fails to accurately identify the cecum and rectum endpoints. This occurs when the colon has sharp bends, resulting in incorrect centreline generation due to faulty skeletonisation. Since the segmentation does not distinguish between the colon wall and lumen, endpoint detection becomes inaccurate. In cases where the colon forms self-connecting loops in segmentation, endpoint identification becomes unreliable. Separately identifying the lumen and colon wall could further enhance the robustness of the proposed method. 4 Conclusion This study presented a three-stage, topology-preserving pipeline for colon segmentation from CT images. The proposed framework addresses a common limitation in deep learning-based colon segmentation, where disconnected segments often occur due to complex colon geometry and limited topological awareness in conventional models. By integrating centreline-based bridging and reconstruction, the pipeline effectively restores continuity while maintaining segmentation accuracy. Experimental results on the TS and RAOS datasets show consistent improvements in topology-related metrics, indicating enhanced anatomical coherence of colon segmentations. Although a minor decrease in overlap-based metrics is observed in some instances, future work will evaluate whether this reduction is acceptable by assessing performance on downstream tasks, such as region-of-interest extraction for colorectal cancer and diverticular disease. Further work will also investigate learning-based centreline alignment to enhance segmentation accuracy. 5 Acknowledgements This research was supported by computational resources provided by the National Computational Infrastructure (NCI), Australia, and Katana, the High Performance Computing facility at the University of New South Wales. 6 Human and/or Animal Subjects This study was conducted using publicly available datasets, and therefore did not require ethical approval. References [1] I. Bakurov, M. Buzzelli, R. Schettini, M. Castelli, and L. Vanneschi (2022) Structural similarity index (SSIM) revisited: A data-driven approach. Expert Systems with Applications 189, p. 116087 (en). External Links: ISSN 09574174, Link, Document Cited by: §2.5. [2] A. Bert, I. Dmitriev, S. Agliozzo, N. Pietrosemoli, M. Mandelkern, T. Gallo, and D. Regge (2009) An automatic method for colon segmentation in CT colonography. Computerized Medical Imaging and Graphics 33 (4), p. 325–331 (en). External Links: ISSN 08956111, Link, Document Cited by: §1. [3] Y. Gao, Y. Jiang, Y. Peng, F. Yuan, X. Zhang, and J. Wang (2025) Medical Image Segmentation: A Comprehensive Review of Deep Learning-Based Methods. Tomography 11 (5), p. 52 (en). External Links: ISSN 2379-139X, Link, Document Cited by: §2.2. [4] K. Gayathri Devi and R. Radhakrishnan (2015) Automatic Segmentation of Colon in 3D CT Images and Removal of Opacified Fluid Using Cascade Feed Forward Neural Network. Computational and Mathematical Methods in Medicine 2015, p. 1–15 (en). External Links: ISSN 1748-670X, 1748-6718, Link, Document Cited by: §1. [5] X. Luo, Z. Li, S. Zhang, W. Liao, and G. Wang (2024) Rethinking Abdominal Organ Segmentation (RAOS) in the Clinical Scenario: A Robustness Evaluation Benchmark with Challenging Cases. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2024, Vol. 15009, p. 531–541 (en). External Links: ISBN 978-3-031-72113-7 978-3-031-72114-4, Link, Document Cited by: §2.5. [6] J. Ma, F. Li, and B. Wang (2024) U-Mamba: Enhancing Long-range Dependency for Biomedical Image Segmentation. arXiv. Note: Version Number: 1 External Links: Link, Document Cited by: §2.2. [7] C. Metz, M. Schaap, S. Klein, P. Rijnbeek, L. Neefjes, N. Mollet, C. Schultz, P. Serruys, W. Niessen, and T. Van Walsum (2012) Alignment of 4D Coronary CTA with Monoplane X-ray Angiography. In Augmented Environments for Computer-Assisted Interventions, Vol. 7264, p. 106–116 (en). Note: Series Title: Lecture Notes in Computer Science External Links: ISBN 978-3-642-32629-5 978-3-642-32630-1, Link, Document Cited by: §2.5. [8] D. Müller, I. Soto-Rey, and F. Kramer (2022) Towards a guideline for evaluation metrics in medical image segmentation. BMC Res Notes 15 (1), p. 210 (en). External Links: ISSN 1756-0500, Link, Document Cited by: §2.5. [9] M. A. Rahman, S. Singh, K. Shanmugalingam, S. Iyer, A. Blair, P. Ravindran, and A. Sowmya (2023) Attention and Pooling based Sigmoid Colon Segmentation in 3D CT images. In 2023 International Conference on Digital Image Computing: Techniques and Applications (DICTA), Port Macquarie, Australia, p. 312–319. External Links: ISBN 979-8-3503-8220-4, Link, Document Cited by: §1. [10] S. Shit, J. C. Paetzold, A. Sekuboyina, I. Ezhov, A. Unger, A. Zhylka, J. P. W. Pluim, U. Bauer, and B. H. Menze (2021) clDice - a Novel Topology-Preserving Loss Function for Tubular Structure Segmentation. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Nashville, TN, USA, p. 16555–16564. External Links: ISBN 978-1-6654-4509-2, Link, Document Cited by: §2.5. [11] L. L. Strate and A. M. Morris (2019) Epidemiology, Pathophysiology, and Treatment of Diverticulitis. Gastroenterology 156 (5), p. 1282–1298.e1 (en). External Links: ISSN 00165085, Link, Document Cited by: §1. [12] C. Wang, Z. Cui, J. Yang, M. Han, G. Carneiro, and D. Shen (2023) BowelNet: Joint Semantic-Geometric Ensemble Learning for Bowel Segmentation From Both Partially and Fully Labeled CT Images. IEEE Trans. Med. Imaging 42 (4), p. 1225–1236 (en). External Links: ISSN 0278-0062, 1558-254X, Link, Document Cited by: §1, §2.2. [13] J. Wasserthal, H. Breit, M. T. Meyer, M. Pradella, D. Hinck, A. W. Sauter, T. Heye, D. T. Boll, J. Cyriac, S. Yang, M. Bach, and M. Segeroth (2023) TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images. Radiology: Artificial Intelligence 5 (5), p. e230024 (en). External Links: ISSN 2638-6100, Link, Document Cited by: §1, §2.5. [14] W. Wu, R. Lei, K. Niu, R. Yang, and Z. He (2022) Automatic segmentation of colon, small intestine, and duodenum based on scale attention network. Medical Physics 49 (11), p. 7316–7326 (en). External Links: ISSN 0094-2405, 2473-4209, Link, Document Cited by: §1. [15] X. Zhang, Z. Xu, L. Shang, Q. Yang, H. Ye, H. Liu, Y. Zou, Y. Lu, Z. Zheng, M. Li, P. Wang, and J. Zhu (2025) Global burden of colorectal cancer attributable to metabolic risks from 1990 to 2021, with predictions to 2046. BMC Cancer 25 (1), p. 228 (en). External Links: ISSN 1471-2407, Link, Document Cited by: §1.