Georgia Institute of Technology
Lightweight Transformers for Echocardiographic Ejection Fraction Estimation: Frame Encoding Strategies and Anatomical Auxiliary Supervision
Abstract
dc:description.abstractAutomated estimation of left ventricular ejection fraction from echocardiographic video could reduce the time and variability associated with manual assessment. This thesis investigates how frame-encoding strategies and anatomical auxiliary supervision affect a lightweight six-layer transformer trained on 10,030 videos from EchoNet-Dynamic. Four frame encoders were evaluated: a linear projection, an ImageNet-pretrained ResNet-50, an adjacent-frame contrastive encoder, and a same-video contrastive encoder. A separately trained U-Net achieved a validation Dice score of 0.936 on expert-annotated end-diastolic and end-systolic frames and generated ventricular cavity masks, normalized cavity areas, and frame-to-frame area changes as auxiliary training targets. Among the evaluated encoders, the linear projection provided the strongest and most consistent regression performance. It also supported the strongest alignment with the generated anatomical targets, with auxiliary mask Dice scores of 0.88 to 0.91 and median area-trajectory correlations of 0.93 to 0.95 across more than 1,100 eligible test videos. However, stronger anatomical alignment did not produce a consistent or practically meaningful improvement in ejection fraction accuracy. These findings show that frame-encoder choice was the primary performance driver in the evaluated framework and that improved anatomical representation does not necessarily translate into improved clinical regression.
Degree
thesis:*- Name thesis:degree_name
- Biomedical Engineering, MS
- Department thesis:degree_department
- Biomedical Engineering (Joint GT/Emory Department)
- Grantor
- Georgia Institute of Technology
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Wang, Shaunna
- Advisor dc:contributor.advisor
-
- Wang, May Dongmei
- Committee members dc:contributor.committeemember
-
- Coskun, Ahmet
- Dasi, Lakshmi
- ORCID person:identifier.orcid
Rights
- Language dc:language.iso
- English
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/82200