A Causality-Inspired Model for Intima-Media Thickening Assessment in Ultrasound Videos

TL;DR

This work tackles frame-wise IMT assessment in carotid ultrasound videos, where cross-frame style shifts create spurious correlations that can mislead thickening detection. It introduces a causality-inspired framework built on a Structural Causal Model (SCM) to block noncausal backdoor paths $S \dashrightarrow Y$ and strengthen the causal chain $C \rightarrow F \rightarrow Y$ through three modules: Spurious Correlation Elimination (SCE), Causal Equivalence Consolidation (CEC), and Causal Transition Augmentation (CTA). SCE perturbs auxiliary style statistics with AdaIN and enforces prediction invariance; CEC performs content randomization with adversarial training to emphasize content-driven predictions; CTA adds a text-based auxiliary pathway guided by chain-of-thought and aligns it with image features via contrastive learning. On an in-house dataset of 120 carotid ultrasound videos, the approach achieves $86.93\%$ accuracy and outperforms several baselines, demonstrating the practical value of incorporating causal analysis and multimodal prompts for robust frame-wise IMT assessment.

Abstract

Carotid atherosclerosis represents a significant health risk, with its early diagnosis primarily dependent on ultrasound-based assessments of carotid intima-media thickening. However, during carotid ultrasound screening, significant view variations cause style shifts, impairing content cues related to thickening, such as lumen anatomy, which introduces spurious correlations that hinder assessment. Therefore, we propose a novel causal-inspired method for assessing carotid intima-media thickening in frame-wise ultrasound videos, which focuses on two aspects: eliminating spurious correlations caused by style and enhancing causal content correlations. Specifically, we introduce a novel Spurious Correlation Elimination (SCE) module to remove non-causal style effects by enforcing prediction invariance with style perturbations. Simultaneously, we propose a Causal Equivalence Consolidation (CEC) module to strengthen causal content correlation through adversarial optimization during content randomization. Simultaneously, we design a Causal Transition Augmentation (CTA) module to ensure smooth causal flow by integrating an auxiliary pathway with text prompts and connecting it through contrastive learning. The experimental results on our in-house carotid ultrasound video dataset achieved an accuracy of 86.93\%, demonstrating the superior performance of the proposed method. Code is available at \href{https://github.com/xielaobanyy/causal-imt}{https://github.com/xielaobanyy/causal-imt}.

Figures (5)

• Figure 1: The causal mechanism for intima-media assessment: (a) Cross-frame style shifts due to view variations of scanning; (b) Structural Causal Model (SCM).
• Figure 2: The architecture of the proposed method. Specifically, we apply Spurious Correlation Elimination (SCE) module to remove non-causal style effects by enforcing prediction invariance with style perturbations (Sect. \ref{['SCE']}), then we design Causal Equivalence Consolidation (CEC) module to strengthen causal content correlation by adversarial optimization during content randomization (Sect. \ref{['CEC']}). Furthermore, we design a Causal Transition Augmentation (CTA) module to enhance causal effects by integrating an auxiliary causal pathway and linking it by contrastive learning (Sect. \ref{['CTA']}).
• Figure 3: Generation of intima-media symptom descriptions using step-type prompts.
• Figure 4: Grad-CAM visualization results for different methods. The first row shows the original video frames, with the first four representing non-thickened frames and the last four representing thickened frames.
• Figure 5: t-SNE visualization in a batch with and without intervention.