Deep Learning Approaches for the Segmentation and Reconstruction of the Hepatic Vascular and Biliary Trees

Clinical context and objectives

This research internship will launch a new collaboration between the Creatis laboratory and the Hospices Civils de Lyon (HCL), with the possibility of continuing the work as a PhD project. The project aims to develop innovative image-analysis methods based on recent advances in artificial intelligence and deep learning, in order to improve the 3D modelling of hepatic vascular structures (the portal venous system, hepatic veins and hepatic arteries) and the biliary tree. The goal is to provide reliable, accurate tools that can be integrated into clinical practice, particularly in multiphase contrast-enhanced CT and biliary MRI (MRCP) protocols, to provide patient-specific visualisation of hepatic anatomy. These tools should facilitate preoperative planning for liver surgery and transplantation, and support clinical decision-making by reducing clinicians’ workload and improving the reproducibility of assessments.

Methodological context and objectives

In recent years, automatic segmentation of the liver and its internal structures has benefited greatly from advances in deep learning. UNet architectures [1] and their more recent variants [2] now underpin most approaches. They achieve strong performance in segmenting the liver parenchyma, enabling reliable volume estimation for preoperative planning [3, 4]. By contrast, segmenting hepatic vascular structures remains more challenging because of their branching, interconnected organisation and their inter-individual variability, particularly in the presence of cirrhosis or large tumours. As a result, segmentation performance varies more widely across studies [4, 5]. The challenge is even greater for the biliary tree: its low contrast on CT images often makes magnetic resonance cholangiopancreatography (MRCP) necessary.
Most existing studies focus on a single vascular or biliary tree in a single imaging modality [4–7]. To address the clinical need, our project aims to segment and reconstruct all vascular and biliary trees using multiphase, multimodal imaging (liver CT and biliary MRI), for structures ranging in size from less than 2 mm (bile ducts) to more than 1 cm (the vena cava). The main challenge is to preserve the connectivity, continuity and topology of these trees despite acquisition noise, partial-volume effects and low contrast. A few recent studies have addressed continuity and topological consistency in the segmentation of hepatic veins [7] and bile ducts [6].
A major obstacle to supervised learning is the lack of annotated image datasets that provide reference labels for all the target structures. Indeed, available public datasets [8] often provide annotations for only one vascular tree, combining several vessel types.

Research programme

Following a literature review of vessel segmentation and reconstruction methods, the student will become familiar with liver images using available public datasets. These include VSNet [7], which comprises approximately 300 images with annotations of the hepatic veins and the portal vein, and LiVS [8], which provides sparse portal-vein annotations for approximately 500 images. These data will be supplemented with images provided by the HCL during the project. To become familiar with the target hepatic structures, the student will also explore interactive segmentation methods using 3D Slicer and its vessel-segmentation plugins, such as RVesselX and VSMOD.
Deep-learning methods for automatic segmentation will then be studied for a selected imaging modality and target structure or structures. This is the main methodological objective of the research internship. The first step will be to implement and evaluate a standard UNet for supervised segmentation [9]. The specific features of recent methods [6, 7] will be examined to determine whether topological constraints [10, 11], which enforce global connectivity of the vascular tree, should be introduced alongside local constraints [12], which model vessel geometry in the neighbourhood of each point. Reconstructing complete individual vascular trees will likely require analysis and post-processing of the predicted segmentations, which may contain breaks and false positives. This reconstruction could make use of the geometric properties of each structure as prior information.

Candidate profile

Applicants should hold a first-year Master’s degree (M1), or be currently enrolled in the second year of a research-oriented Master’s programme (M2), with a strong background in applied mathematics, image analysis and computer science, as well as good programming skills. Theoretical and practical knowledge of deep-learning methods is required. Familiarity with medical imaging would be an advantage.

Application

Please email your CV and recent academic transcripts.

Internship allowance

Approximately €640 per month, in accordance with current French regulations.

References

  1. Çiçek Ö, et al. 3D U-Net: learning dense volumetric segmentation from sparse annotation. MICCAI. 2016.
  2. Xia Q, et al. A comprehensive review of deep learning for medical image segmentation. Neurocomputing. 2025.
  3. Wei D, et al. A review of advancements and challenges in liver segmentation. Med Image Anal. 2024.
  4. Sethia K, et al. Advances in liver, liver lesion, hepatic vasculature, and biliary segmentation. Artif Intell Rev. 2025.
  5. Song Z, et al. Liver vessel segmentation using deep learning. Heliyon. 2024.
  6. Zhou G, et al. Tagnet: A transformer-based axial guided network for bile duct segmentation. Biomedical Signal Processing and Control. 2023.
  7. Xu J, et al. VSNet: Vessel Structure-aware Network for hepatic and portal vein segmentation. Medical Image Analysis. 2025.
  8. Gao Z, et al. Laplacian salience-gated feature pyramid network for accurate liver vessel segmentation. IEEE Trans Med Imag. 2023.
  9. Isensee F, et al. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods. 2021;18(2):203–211.
  10. Shit S, et al. clDice—a novel topology-preserving loss function for tubular structure segmentation. IEEE CVF. 2021.
  11. Kirchhoff Y, et al. Skeleton recall loss for connectivity conserving and resource efficient segmentation of thin tubular structures. ECCV. 2024.
  12. Kervadec H, et al. Boundary loss for highly unbalanced segmentation. Med Image Anal. 2021.