| MS Seminar

Name of the Speaker: Ms. Shital Sheshrao Yelne (EE20S052)
Guide: Prof. Mohanasankar S
Online meeting link: https://meet.google.com/sjj-gcvi-vup
Date/Time: January 19th 2024 (Friday), 3:00 PM
Title: Adapting Deep Learning Models for Diverse Medical Imaging Data: Cross- Scanner and Cross-Modality Challenges

Abstract

In In the field of ultrasound image interpretation, the unique texture of ultrasound scanners becomes a distinguishing feature, especially in anatomical locations such as the liver. Deep Learning (DL) models trained on certain scanner type images not only capture anatomical content but also learn the scanner’s inherent texture, which serves as a type of anatomical marker. However, learnt styles impede the portability of such models across multiple scanner types, resulting in poor results, particularly in segmentation models, where lower Dice values are observed when applied to images from various scanner types. A novel strategy, rather than retraining the DL model, includes changing the texture of previously unseen data to fit with the training distribution. Leveraging features from a pre-trained DL model, such as VGG network, enables Neural Style Transfer, reducing the algorithmic complexity. This methodology demonstrates improved segmentation outcomes without requiring the retraining of existing models. This need for adaptability extends to MRI interpretation.

Shifting focus to MRI image interpretation, a critical aspect in medical imaging diagnostics, the challenge lies in obtaining multiple versions of the imaging to enhance diagnostic accuracy. Deep learning, specifically through style transfer using CycleGAN, provides a solution by generating artificial MRI images with varying contrast levels from existing scans. This approach, employing Cycle Generative Adversarial Networks (CycleGANs), facilitates the creation of T2-weighted images from T1-weighted MRI images and vice versa. The CycleGAN model is trained to understand the intricate relationships between T1 and T2 weighted images, enabling seamless transformations between the two contrasts. The technique employed in our work can be aptly characterized as cross-modality style transfer.