| MS Seminar

Name of the Speaker: Ms. Tanvi Kulkarni (EE20S046)
Guide: Prof. Mohanasankar S
Online meeting link: https://meet.google.com/aoi-tnyx-hma
Date/Time: 5th June 2024 (Wednesday), 2:15 PM
Title: Learning to Atlas Register for Rapid Segmentation of Brain Structures and Improved Deep Similarity Metrics

Abstract :

Registration is one of the most preliminary steps in many medical imaging downstream tasks. The registration process essentially aligns images to capture the same overlapping anatomical structures in the input images. Recently, Deep Learning (DL) algorithms have been shown to deliver competitive registration performance as compared to traditional iterative algorithms with a significant computational speedup and efficiency. However, these DL methods can fail to perform for multimodal datasets when the input images display different intensity characteristics, for example, MRI T1 to T2 contrast. DL-based registration is also used to derive segmentation maps, popularly known as atlas-based segmentation (ABS), to segment multiple structures in a single operation. Such methods are often applied for diagnostic and surgical planning purposes. However, due to the training data’s limited availability and restricting nature, DL-based networks can produce incorrect segmentation maps. Therefore, an automatic quality control tool is necessary to identify such failed ABS processes and take corrective action before proceeding to further downstream tasks, especially when accurate segmentation maps are paramount for safe healthcare applications.

In this seminar, we aim to present the steps taken to optimize multiple stages of the DL registration pipeline, which can be critical for efficient process improvements in clinical applications. More specifically, we first study the application of a state-of-the-art learning-based deformable registration framework, VoxelMorph, for challenging Magnetic Resonance Imaging (MRI) datasets of fetal and adult human brains and sub-micron level Nissl-stained histology datasets of adult mouse brain. Secondly, we explore the advantages of leveraging intensity and texture-based features combined with the shape features of the predicted segmentation map to develop a Machine Learning (ML) model for quality control of ABS without ground-truth segmentation maps. Thirdly, we propose a novel data-driven DL-based similarity metric, SiamRegQC. SiamRegQC leverages the semantic representations of the source and target images obtained by training a Siamese network on a two-step binary misalignment classification task. We demonstrate that SiamRegQC can provide better registration performances than traditional and previously proposed deep similarity metrics with a maximum SSIM score of 0.825 for the multi-modal deformable registration between MRI T1 and T2 images.