Please note: This PhD seminar will take place in DC 3317.
Blake VanBerlo, PhD candidate
David R. Cheriton School of Computer Science
Supervisors: Professors Jesse Hoey, Alexander Wong
A central component of joint embedding self-supervised learning is the choice of data augmentation pipeline used to produce positive pairs. This study investigated whether stochastic transformations designed to preserve semantic content in ultrasound images would improve self-supervised learning methods applied to diagnostic tasks in ultrasound.
Three pipelines were investigated: (1) a baseline pipeline commonly used in different modalities, (2) AugUS-v1 — a pipeline composed of novel and standard transformations designed to retain semantic content, and (3) AugUS-v2 — a pipeline composed from transformations in the baseline and AugUS-v1 pipelines that were observed to improve validation set performance. The results indicated that semantics-preserving transformations can improve the performance of self-supervised models for ultrasound for some tasks, but that other transformations that are likely to produce semantically inconsistent pairs may be required in addition to achieve top performance on other tasks.