Gaze-Assisted Human-Centric Domain Adaptation for Cardiac Ultrasound Image Segmentation

TL;DR

Cardiac ultrasound segmentation suffers from domain shifts across devices and acquisition settings. The authors introduce gaze-assisted human-centric domain adaptation (GAHCDA), a teacher–student framework that leverages doctor gaze heatmaps through two modules: Gaze Augment Align (GAA), which fuses gaze cues with features via cross-attention to extract human-cognition general features, and Gaze Balance Loss (GBL), which weights learning toward gaze-dense regions to curb pseudo-label noise. On CAMUS as the source and HMC-QU as the target, GAHCDA delivers state-of-the-art performance (DSC $=76.14\%$, ASSD $=6.976$) and outperforms GAN-based and self-training baselines, with ablations confirming the contribution of each module. The approach demonstrates the practical potential of doctor gaze guidance to enhance cross-domain ultrasound segmentation, reduce annotation needs, and improve clinical applicability.

Abstract

Domain adaptation (DA) for cardiac ultrasound image segmentation is clinically significant and valuable. However, previous domain adaptation methods are prone to be affected by the incomplete pseudo-label and low-quality target to source images. Human-centric domain adaptation has great advantages of human cognitive guidance to help model adapt to target domain and reduce reliance on labels. Doctor gaze trajectories contains a large amount of cross-domain human guidance. To leverage gaze information and human cognition for guiding domain adaptation, we propose gaze-assisted human-centric domain adaptation (GAHCDA), which reliably guides the domain adaptation of cardiac ultrasound images. GAHCDA includes following modules: (1) Gaze Augment Alignment (GAA): GAA enables the model to obtain human cognition general features to recognize segmentation target in different domain of cardiac ultrasound images like humans. (2) Gaze Balance Loss (GBL): GBL fused gaze heatmap with outputs which makes the segmentation result structurally closer to the target domain. The experimental results illustrate that our proposed framework is able to segment cardiac ultrasound images more effectively in the target domain than GAN-based methods and other self-train based methods, showing great potential in clinical application.

Figures (8)

• Figure 1: (a) Domain adaptation based on GAN is prone to mode collapse. (b) Domain adaptation based on self-training is prone to overfit on pseudo labels. (c) Human-centric domain adaptation uses gaze information from doctors’ diagnoses to help the model adapt to the domain and reduce overfitting.
• Figure 2: The Gaze-Assisted Human-Centric domain adaptation method uses gaze guidance to help student models acquire human-like cross-domain segmentation and recognition capabilities. (a) Teacher model parameters trained on the source domain are used to initialize the student model in the target domain. (b) The Gaze Augment Align Module (GAA) module uses gaze information to help the model obtain human cognition general features. (c) The Gaze Balance Loss (GBL) helps the model solve the over/under-segmentation problem.
• Figure 3: The GAA helps the student model leverage general features to acquire the ability to recognize and segment regions across domains.
• Figure 4: Visual comparison of outputs. The results show that our GAHCDA uses gaze information to obtain results closer to GT.
• Figure 5: The ablation study demonstrated that GAA and GBL effectively eliminate domain differences by leveraging human cognition, thereby improving the model's segmentation performance in the target domain. Ablation study of (a) GAHCDA. (b) Hyper-parameter $\lambda_{gaa}$. (c) Hyper-parameter $\lambda_{gb}$.