Improved Representation Learning with Multitasking in Blind Sweep Obsetric Ultrasound Videos

Nishant Luitel, Kanchan Poudel, Prasiddha Bhandari, Bishram Acharya, Angelina Ghimire, Pradeep Raj Regmi, Bishesh Khanal

Keywords: Blind Sweep, Fetal, Multi-task, Multi-head, Contrastive

TL;DR: Improved Representation for Blind Sweep Obstetric Ultrasound Videos with multitask learning

Abstract:

Blind Sweep Obstetric Ultrasound (BSOU), based on predefined abdominal trajectories, enable non-experts to capture ultrasound videos and are increasingly used for AI-based estimation of obstetric measures like Gestational Age and Fetal Presentation. However, existing work focuses on single-task models, overlooking the potential of joint learning.
We propose the first multi-task framework for BSOU-based AI models, leveraging spatio-temporal constraints inherent in sweep protocols and fetal anatomy.
Our approach includes multi-head cross-entropy (MTCE) and a novel approach to Multi-head Supervised Contrastive Loss (MTSCon) for BSOU datasets, treating videos with matching labels across patients and sweep types as augmented versions of the same input in a contrastive setting.
We introduce new applications, Deepest Vertical Pocket estimation and sweep type prediction, and show that carefully selected task combinations improve both in-domain performance and generalization to out-of-domain settings, with MTSCon offering better gains in representation quality.

Primary Subject Area: Unsupervised Learning and Representation Learning

Secondary Subject Area: Application: Radiology

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 269