Convolutional Occupancy Networks for Medical Imaging with Applications to the KiTS23 Challenge

Lukasz Nowakowski, Rajni Patel

Confirmation: I have read and agree with the IEEE BHI 2025 conference submission's policy on behalf of myself and my co-authors.

Keywords: Occupancy Networks, Medical Imaging, Segmentation, KiTS23 Challenge

TL;DR: We propose an application of occupancy networks for 3D medical image segmentation, demonstrating their effectiveness on the publicly available KiTS23 dataset.

Abstract: We propose an application of occupancy networks for 3D medical image segmentation, demonstrating their effectiveness on the publicly available KiTS23 dataset. Unlike conventional CNN-based methods that operate in voxel space using encoder-decoder architectures, our approach represents anatomical structures as continuous decision boundaries within normalized coordinate space. This formulation enables fine-grained surface delineation and flexible inference resolution. Our architecture integrates a MedicalNet-pretrained ResNet encoder, a multi-scale Bi-directional Feature Pyramid Network (BiFPN) feature fusion backbone, and class-specific parallel prediction heads. To address the high anatomical variability and class imbalance in the dataset, we design a training strategy based on structured 3D patch sampling, coupled with a targeted refinement mechanism during inference that leverages coarse predictions to guide high-resolution queries for underrepresented classes. Extensive experiments show that our model achieves competitive performance on Dice and Surface Dice metrics compared to leaderboard methods. These results underscore the potential of continuous occupancy-based representations for high-fidelity medical segmentation.

Track: 7. General Track

Registration Id: 3KNHN2QZPWP

Submission Number: 246