Towards clinical application of liver, vessel, and tumor segmentation using partially labeled data

Eirik Agnalt Østmo, Keyur Radiya, Kristoffer Knutsen Wickstrøm, Michael Kampffmeyer, Karl Øyvind Mikalsen, Robert Jenssen

Keywords: Medical image segmentation, computed tomography, contrast-enhanced CT, liver vessel tumor segmentation, multi-label segmentation, partial-label learning

TL;DR: Investigating multi-label masked loss strategies to learn liver, vessel, and tumor segmentation from partially labeled CT datasets and demonstrating clinical utility of the application.

Abstract: Accurate delineation of liver parenchyma, intrahepatic vessels, and tumors (LVT) may aid earlier tumor detection, consistent response assessment, and surgical planning for patients with liver cancer. Deep learning (DL) may enable such automated delineation, but available CT datasets are fragmented and partially labeled, making them unsuited for end-to-end training. We investigate a single-head, 3D segmentation framework that learns from such fragmented data by: (i) loss masking per class or voxel to ignore missing annotations, (ii) using multi-hot targets and the anatomical hierarchy inherent to liver, vessels, and tumors, to handle overlapping structures without class competition. In controlled ablations that simulate partial-label training, this multi-label masked strategy reliably outperforms masked multi-class baselines, avoids precision collapse, and improves tumor overlap and lesion detection sensitivity. Scaling training to multiple partially labeled datasets, the model surpasses full-resolution nnU-Net on an external clinical cohort, with higher tumor and vessel segmentation performance. We conduct a qualitative retrospective case study to illustrate the clinical potential of the LVT application. We find that LVT models can enable earlier detection of metastasis by six months, longitudinal size tracking aligned with radiologist measurements, 3D tumor–vessel visualization for surgical planning, and stable inter-phase liver volumetry (~2\% deviation). These results show that multi-label masked learning enables robust, clinically relevant LVT segmentation from partially labeled datasets.

Serve As Reviewer: ~Karl_Øyvind_Mikalsen1, ~Kristoffer_Knutsen_Wickstrøm1

Submission Number: 44