Source-linked AI summary
Preserving Anatomical Continuity: Three-Stage Pipeline for Colon Segmentation in 3D Abdominal CT Scans
Deshan Kalupahana, Sonit Singh, Praveen Ravindran, Arcot Sowmya
TL;DR
CT-based colon segmentation is important for colorectal disease analysis, but deep-learning predictions can be discontinuous. The paper combines segmentation, centreline bridging, and reconstruction, reporting improved topology-related metrics across TS and RAOS, with some overlap decreases.
Problem
Deep-learning colon segmentation can produce disconnected predictions despite the importance of accurate CT segmentation for colorectal disease analysis.
Source: §1 · p. 1Method
A three-stage pipeline performs initial segmentation, centreline bridging, and reconstruction to restore continuity across disconnected regions.
Source: §1 · p. 1, §2.3 · p. 2, §4 · p. 4Results
Across TotalSegmentator and RAOS, topology-related metrics improved consistently, although some instances showed minor decreases in overlap-based metrics.
Source: §4 · p. 4Takeaways & Limitations
The reported topology improvements indicate more anatomically coherent colon segmentations on the evaluated datasets.
Source: §4 · p. 4Quantitative gains remain small because reconnected regions are relatively small.
Source: §3.2 · p. 3
Abstract
from arXiv · showhide
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.
1. INTRODUCTION
CT is the primary modality for detecting colorectal disease, making accurate colon segmentation an important preliminary step. This study targets discontinuous deep-learning predictions with a topology-preserving pipeline.
1. INTRODUCTION
- Accurate colon segmentation is a preliminary step in identifying abnormalities associated with colorectal cancer and diverticular disease using CT.The passage describes CT as the primary modality for detecting these diseases.Source: §1 · p. 1
- Earlier thresholding on cleansed, contrast-enhanced bowel scans gave way to deep-learning methods addressing partial-volume effects.Source: §1 · p. 1
- BowelNet achieved approximately 81% average Dice for colon segmentation on a private dataset, while TotalSegmentator reported 89.6% accuracy and still showed discontinuous predictions.Source: §1 · p. 1
- The study introduces a three-stage pipeline intended to restore structural continuity across disconnected areas while maintaining segmentation accuracy.Source: §1 · p. 1
2. METHOD
The method combines an initial segmentation network, topology-driven centreline bridging, and learned reconstruction of disconnected regions. Evaluation uses two CT datasets and overlap, distance, and topology-based measures.
2.1. Overview of the Segmentation Pipeline
- The pipeline performs initial segmentation, bridges centreline gaps, then predicts and merges disconnected regions into a complete, topologically consistent colon structure.Source: §2.1 · p. 1, Figure 1 · p. 2
2.2. Stage 1 : Initial Colon Segmentation
- Stage 1 assesses UNet-based, transformer, and state-space segmentation architectures, including nnUNet, BowelNet, and U-Mamba, using combined cross-entropy and Dice losses.Source: §2.2 · p. 1
2.3. Stage 2 : Centreline Bridging
- Stage 2 skeletonizes predictions, converts them to graphs, removes short branches, and uses anatomical endpoint cues and Bézier curves to bridge disconnected components.The algorithm searches from the rectum endpoint toward the cecum; only centreline bridges, including extended points, are forwarded to reconstruction.Source: §2.3 · p. 2
2.4. Stage 3 : Reconstruction of Full Colon
- Stage 3 uses a UNet trained on CT volumes and ground-truth centrelines to predict bridge regions, filters them by endpoint connectivity, and merges them with the initial segmentation.Source: §2.4 · p. 2
2.5. Experimental Setup
- The evaluation uses TS and RAOS abdominal CT datasets, five-fold cross-validation with identical splits, and manual review of Stage 1 predictions.TS was manually selected from TotalSegmentator and RAOS was manually annotated.Source: §2.5 · p. 3
- Overlap metrics measure volumetric agreement, distance metrics assess boundary deviation and structural similarity, and topology metrics assess continuity and structural consistency.Overlap measures may miss small disconnections that can disrupt anatomical correctness.Source: §2.5 · p. 3
3. RESULTS AND ANALYSIS
The pipeline’s final reconstruction improves topology in targeted cases, but quantitative gains are modest and visual continuity may be clearer than metrics suggest.
3.1. Initial Colon Segmentation
- nnUNet performed best on both datasets; close GT alignment was TS 49.25% and RAOS 52.78%, while disconnections affected TS 18.41% and RAOS 23.48%.Some predictions were visually more accurate than GT: TS 9.20% and RAOS 18.40%.Source: §3.1 · p. 3
3.2. Final Colon Reconstruction
- Across all samples, Stage 3 metrics decreased slightly in cases where predictions overpredicted or reconstructed regions missing from disconnected GT annotations.The authors note that reconstruction can successfully restore regions absent from the GT annotations.Source: §3.2 · p. 3
- The initially disconnected-prediction group had lower overlap metrics than the full-sample group, and GT annotations were disconnected in 40 TS and 50 RAOS cases.The authors say these GT disconnections can obscure quantitative improvement.Source: §3.2 · p. 3
- Topology-based metrics improve under distance thresholds, while overlap and distance metrics can dip, especially for large gaps; gains are strongest below 90mm.The thresholds connect endpoints less than 120mm apart and, in the refined setting, less than 90mm apart.Source: §3.2 · p. 3, Table 2 · p. 4, Table 3 · p. 4
- Quantitative improvements remained small because the reconnected regions were relatively small and contributed minimally to overall metric gains.Source: §3.2 · p. 3
3.3. Qualitative analysis
- Visual inspection showed improved continuity and reconnection in both datasets, which was not fully reflected by quantitative metrics.The figure presents disconnections and progressive reconstruction across pipeline stages against GT annotations.Source: §3.3 · p. 4, Figure 2 · p. 4
- Sharp bends and self-connecting loops can make centreline generation or endpoint identification unreliable when segmentation does not distinguish colon wall from lumen.The authors suggest separate lumen and wall segmentation could improve robustness.Source: §3.3 · p. 4
4. CONCLUSION
The study presents a three-stage, topology-preserving pipeline for colon segmentation and reports improved topology-related metrics on two datasets. Overlap metrics decrease slightly in some cases, and downstream evaluation remains future work.
4. CONCLUSION
- The pipeline combines centreline-based bridging and reconstruction to restore continuity in deep learning colon segmentations.Source: §4 · p. 4
- Experiments on TS and RAOS show consistent improvements in topology-related metrics, with minor overlap-based metric decreases in some instances.Source: §4 · p. 4
- Future work will test whether overlap decreases are acceptable for downstream region-of-interest extraction and investigate learning-based centreline alignment.Source: §4 · p. 4
6. HUMAN AND/OR ANIMAL SUBJECTS
The study used publicly available datasets and did not require ethical approval.
6. HUMAN AND/OR ANIMAL SUBJECTS
- The study used publicly available datasets and therefore did not require ethical approval.Source: §6 · p. 5