Diffusion Model Based Knee Cartilage Segmentation in MRI

MRI imaging is crucial for knee joint analysis in osteoarthritis (OA) diagnosis. The segmentation and thickness estimation of knee cartilage are vital steps for OA assessment. Most deep learning algorithms typically produce a single segmentation mask or rely on architectural modifications like Dropout to generate multiple outputs. We propose an alternative approach using Denoising Diffusion Models (DDMs) to yield multiple variants of segmentation outputs for knee cartilage segmentation and thus offer a mechanism to study predictive uncertainty in unseen test data. We further propose to integrate sparsity adaptive losses to supervise the diffusion process to handle intricate knee cartilage structures. We could empirically validate that DDM-based models predict more meaningful uncertainties when compared to Dropout based mechanisms. We have also quantitatively shown that DDM-based multiple segmentation generators are resilient to noise and can generalize to unseen data acquisition setups.