Cross-Scanner Domain Generalisation in Echocardiography via Frequency-Based Style Normalisation
Supervisory Team
Introduction
Artificial intelligence has demonstrated strong potential in echocardiography, yet real-world deployment remains limited due to cross-scanner variability. Differences in acquisition settings and manufacturer-specific processing lead to domain shift, reducing model performance on unseen scanners. This PhD project investigates frequency-based style normalisation techniques grounded in Fourier analysis to develop scanner-invariant deep learning models for echocardiography. The approach aims to improve robustness across tasks such as classification and segmentation without requiring target-domain data.
Aim
To develop frequency-domain approaches that improve cross-scanner domain generalisation in echocardiography, enabling robust and scanner-invariant deep learning models.
Objectives
-
Analyse scanner-induced variability in echocardiography and its frequency characteristics.
-
Develop frequency-based style normalisation techniques for domain generalisation.
-
Evaluate performance across classification and segmentation tasks.
-
Benchmark against conventional augmentation and style transfer methods.
-
Validate cross-scanner robustness using multi-domain public datasets.
Methodology
Literature Review and Dataset Analysis:
Comprehensive review of domain generalisation methods and analysis of frequency characteristics across echocardiography datasets.
Frequency-Based Normalisation:
Application of Fourier Transform to decompose images into amplitude and phase spectra, enabling controlled amplitude mixing or normalisation to promote scanner invariance.
Model Development:
Integration of frequency-based techniques into deep learning architectures for classification and segmentation.
Baseline Comparisons:
Evaluation against standard spatial augmentation, histogram matching and adversarial style transfer approaches.
Cross-Domain Evaluation:
Leave-one-domain-out experiments across publicly available datasets to assess zero-shot scanner generalisation.
Timeline
-
Year 1: Literature review, dataset acquisition, statistical and frequency-domain analysis, baseline implementation.
-
Year 2: Development of frequency-based framework, extensive experimentation across tasks, ablation studies and intermediate publications.
-
Year 3: Cross-scanner validation, clinical relevance assessment, dissemination of findings and thesis completion.
Expected Outcomes
-
Development of scanner-invariant deep learning methods for echocardiography.
-
Improved robustness of AI models across unseen scanner vendors.
-
A lightweight and scalable framework suitable for real-world clinical deployment.
-
Theoretical contributions linking frequency characteristics to domain shift in ultrasound imaging.
-
Peer-reviewed publications in medical imaging and computer vision venues.