| MS Seminar

Name of the Speaker: Ms. MUDIYAM VEERASRAVANTHI (EE20S047)
Guide: Prof. Mohanasankar S
Online meeting link: https://meet.google.com/bfw-cskv-foa
Date/Time: 17th January 2024 (Wednesday), 2:30 PM
Title: Knee Cartilage Segmentation and Thickness Estimation Using MRI.

Abstract

Magnetic Resonance Imaging (MRI) is integral to the osteoarthritis (OA) diagnostic process for knee joints. This involves crucial steps like segmenting and estimating the thickness of knee cartilage. Given the structural similarity of cartilage to its surroundings, employing multiple variants of segmentation mask is preferable for comprehensive analysis. Traditional deep learning methods generate a single segmentation mask or utilize techniques like Dropout for generating multiple variants of segmentation masks. In this study, we employed an approach using Denoising Diffusion Models (DDMs) to produce diverse segmentation outputs, allowing for the exploration of predictive uncertainty in unseen data. Additionally, we integrated sparsity adaptive losses to guide the diffusion process, particularly for intricate knee cartilage structures. Our empirical validation substantiates that DDM-based models offer more meaningful uncertainties compared to their Dropout-based counterparts. This evidence underscores the effectiveness of DDMs in enhancing the interpretability and reliability of uncertainty predictions in knee cartilage segmentation. Furthermore, our quantitative analyses underscore the robustness of DDM-based segmentation mask generators. These models exhibit resilience to noise, a critical consideration in real-world imaging scenarios, and demonstrate notable effectiveness in generalizing to previously unseen data acquisition setups. After acquiring segmentation masks, we estimate the cartilage thickness, a crucial parameter linked to joint functionality. Thicker and healthier cartilage contributes to smoother joint movement. Monitoring the changes in cartilage thickness over time enhances the assessment of osteoarthritis progression. We propose a method for cartilage surface separation and calculate the thickness values. These thickness values help in generating 2D and 3D thickness map visualizations. In the presence of significant deformations, visualization of these thickness maps offers valuable insights into the extent of deformation.