<aside> 💡 Summary We developed an automated vessel‑segmentation pipeline for multiphase abdominal CT to assist surgical planning. Using pairs of arterial phase (AP) and portal phase (PP) scans, the system extracts continuous vascular trees, including small peripheral branches, and differentiates arterial and venous structures. We designed a custom loss and post‑processing scheme to reduce discontinuities, achieving consistent vessel tracking across phases. The resulting tool improves surgical visibility of vessel anatomy and supports preoperative planning.

</aside>

vessel.png

Background

Accurate vascular segmentation is essential for gastrectomy and other abdominal surgeries, where surgeons need to identify arteries, veins and their relationship to nearby organs. AP scans emphasize arterial structures, while PP scans are taken slightly later to visualize the portal venous system. Veins often have lower contrast and more irregular shapes than arteries, making PP segmentation more challenging. Moreover, abdominal vessels form thin, tree‑like networks; models must capture long, continuous paths without breaks to preserve clinical relevance

Approach

  1. Dataset & Phase Separation – We collected ~300 CT pairs (AP and PP) and treated the two phases separately. Each phase was preprocessed with intensity normalization and resampling to maintain anatomical fidelity. AP models focused on arterial trees; PP models addressed the more subtle portal venous system
  2. Anatomical Challenges – PP segmentation is harder because portal veins blend into surrounding tissues and show greater morphological variability. Furthermore, vessels near organs (e.g., hepatic veins adjacent to the liver) are critical for surgery but difficult to delineate due to overlapping intensities. Our modelling therefore pays particular attention to organ‑adjacent vessels.
  3. Model & Loss Design – We employed volumetric convolutional networks to learn multi‑scale vessel features. To counter bias toward larger vessels and encourage continuity of thin branches, we designed a composite loss that assigns higher weights to small‑diameter vessels based on their relative volume, similar to class‑adaptive weighting strategies used in organ segmentation. This helps the network optimize for both large arteries and small peripheral veins.
  4. Pre‑ and Post‑processing – Morphological filtering and connected‑component analysis were applied to ground‑truth labels and predictions. Pre‑processing helps the model learn continuous vessel paths, while post‑processing reconnects fragmented predictions and eliminates false positives. Because vascular networks are thin and elongated, maintaining uninterrupted centerlines is critical; these steps improved continuity without overthickening the vessels.
  5. Phase‑specific Training – Separate models for AP and PP allowed us to tailor augmentation, loss weighting and evaluation criteria to the different contrast profiles of arteries and veins. This approach accommodates the low contrast and irregularity of portal veins while still capturing the easier, thicker arterial structures.

Results

Qualitatively, the AP model provided clear arterial trees with minimal breakage, while the PP model achieved balanced segmentation of both thick and thin venous branches. Weighted loss design improved sensitivity to small vessels and reduced discontinuities; even peripheral branches were traced further without sacrificing overall accuracy. Compared with conventional Dice‑based segmentation, our weighted loss produced more uniform performance across vessel sizes and phases, addressing the common bias toward large structures. Post‑processing further mitigated fragmentations, ensuring long, continuous vessel paths, which clinical reviewers found essential for surgical planning.

threshold_vs_dice.png

Abdominal Artery (except Aorta) Segmentation Performance in Dice score for each threshold values

threshold_vs_recall.png

Abdominal Artery (except Aorta) Segmentation Performance in Recall score for each threshold values

Conclusions

By explicitly separating AP and PP workflows and designing a loss and post‑processing scheme that emphasizes continuity of thin vessels, we built a robust vessel segmentation tool for abdominal CT. It delivers comprehensive maps of arterial and portal venous trees, including critical branches near organs. These maps enhance preoperative planning by revealing vascular variations and potential surgical risks. Future work may involve integrating multi‑modal data, improving vein‑artery classification, and exploring explainable AI techniques to better communicate model confidence to surgeons.