Core Concepts
The authors propose a Bayesian learning framework using Hamiltonian Monte Carlo (HMC) to improve uncertainty estimation and segmentation accuracy in cardiac MRI. The approach combines cold posterior sampling with cyclical annealing strategies for efficient training.
Abstract
The study addresses the limitations of Bayesian neural networks (BNNs) for medical image segmentation by introducing a novel Bayesian deep learning framework using HMC. The proposed method, HMC-CP, enhances uncertainty estimation and segmentation accuracy for cardiac MRI images. By exploring the posterior distribution with SGHMC variants, the study demonstrates improved functional diversity compared to existing methods like PHi-Seg and MC-Dropout. The research highlights the importance of temperature control in achieving optimal calibration and segmentation performance, especially in the presence of data augmentation. Additionally, the study investigates the impact of prior strength on model performance, emphasizing the trade-off between prior assumptions and model accuracy.
Stats
For an image with N voxels, let p(ˆyi = c|xi, w) be the predictive probability of voxel xi belonging to class c with C semantic classes in total.
We systematically investigated the effect of cold-posterior in the cardiac MRI segmentation network.
We propose an aggregated confidence score that can detect segmentation failure on the image level.
The ACDC dataset was used for training and validation, consisting of short-axis end-diastolic (ED) and end-systolic (ES) cardiac MRI SSFP cine images.
The proposed SGHMC method was trained for NC = 3 cycles of 333 epochs.
Quotes
"The resulting Bayesian DNN outputs an ensemble segmentation along with the segmentation uncertainty."
"Our results show that the proposed method improves both segmentation accuracy and uncertainty estimation."