AKD-BNN: Adaptive Kernel Dynamic Bayesian Neural Networks for Enhanced Medical Image Segmentation with Uncertainty Estimation


Qianying He (City University of Macau), Xuan Liu (City University of Macau), Wenjian Liu (City University of Macau)
The 35th British Machine Vision Conference
PDFPosterVideo (Right click to download)

Abstract

Medical image segmentation faces significant challenges due to spatial heterogeneity and complex feature variations. Traditional Bayesian Neural Networks (BNNs) combined with Gaussian Processes (GPs) rely on fixed kernel functions, limiting their adaptability to diverse image regions. This paper proposes an Adaptive Kernel Dynamic BNN (AKD-BNN) framework, which enhances segmentation performance and uncertainty estimation through a dynamic kernel learning module and multi-scale feature fusion. The approach includes a dynamic kernel module that adjusts GP kernel parameters based on local image semantics, Transformer-based multi-scale feature extraction, and an uncertainty-guided loss function. Experiments on multiple medical image datasets, including brain tumor (BraTS), cardiac MRI (ACDC), and lung CT (LIDC-IDRI), demonstrate that AKD-BNN outperforms state-of-the-art methods such as BNN-GP and DeepLabV3+. Specifically, AKDBNN improves the Dice score from 0.90 to 0.95 on BraTS and reduces the Expected Calibration Error (ECE) by approximately 20%. The results highlight its superior segmentation consistency in complex regions, such as tumor boundaries, providing more reliable uncertainty maps for clinical applications. This study offers a robust and interpretable solution for medical image analysis, with potential extensions to other high-dimensional tasks.

Citation

@inproceedings{He_2025_BMVC,
author    = {Qianying He and Xuan Liu and Wenjian Liu},
title     = {AKD-BNN: Adaptive Kernel Dynamic Bayesian Neural Networks for Enhanced Medical Image Segmentation with Uncertainty Estimation},
booktitle = {36th British Machine Vision Conference 2025, {BMVC} 2025, Sheffield, UK, November 24-27, 2025},
publisher = {BMVA},
year      = {2025},
url       = {https://bmva-archive.org.uk/bmvc/2025/assets/papers/Paper_1109/paper.pdf}
}


Copyright © 2025 The British Machine Vision Association and Society for Pattern Recognition
The British Machine Vision Conference is organised by The British Machine Vision Association and Society for Pattern Recognition. The Association is a Company limited by guarantee, No.2543446, and a non-profit-making body, registered in England and Wales as Charity No.1002307 (Registered Office: Dept. of Computer Science, Durham University, South Road, Durham, DH1 3LE, UK).

Imprint | Data Protection