Deep Spectral Methods for Unsupervised Ultrasound Image Interpretation

TL;DR

This paper tackles unsupervised segmentation of ultrasound images by introducing a deep spectral segmentation framework that leverages self-supervised transformer features (DINO) to build affinity graphs tailored to ultrasound. It combines spectral clustering with ultrasound-specific patchwise affinities and shape/position priors to produce semantically meaningful tissue segments without labels, followed by a second semantic clustering step guided by mask and positional embeddings. Evaluations on three ultrasound datasets demonstrate competitive segmentation quality and boundary preservation relative to baselines, with code made publicly available. The work highlights a trade-off between fine-grained mask fidelity and label consistency, pointing to future directions in feature-space enhancements and self-training for improved unsupervised ultrasound interpretation.

Abstract

Ultrasound imaging is challenging to interpret due to non-uniform intensities, low contrast, and inherent artifacts, necessitating extensive training for non-specialists. Advanced representation with clear tissue structure separation could greatly assist clinicians in mapping underlying anatomy and distinguishing between tissue layers. Decomposing an image into semantically meaningful segments is mainly achieved using supervised segmentation algorithms. Unsupervised methods are beneficial, as acquiring large labeled datasets is difficult and costly, but despite their advantages, they still need to be explored in ultrasound. This paper proposes a novel unsupervised deep learning strategy tailored to ultrasound to obtain easily interpretable tissue separations. We integrate key concepts from unsupervised deep spectral methods, which combine spectral graph theory with deep learning methods. We utilize self-supervised transformer features for spectral clustering to generate meaningful segments based on ultrasound-specific metrics and shape and positional priors, ensuring semantic consistency across the dataset. We evaluate our unsupervised deep learning strategy on three ultrasound datasets, showcasing qualitative results across anatomical contexts without label requirements. We also conduct a comparative analysis against other clustering algorithms to demonstrate superior segmentation performance, boundary preservation, and label consistency.

Figures (5)

• Figure 1: In our unsupervised semantic segmentation pipeline, ultrasound images undergo preprocessing and dense feature extraction to derive feature affinities. Ultrasound-specific affinities are then calculated using similarity metrics (MI, SSD) and combined with initial affinities for spectral clustering, yielding pseudo masks. Subsequently, images are cropped to focus on detected segments, and dense features alongside positional and shape priors refine clustering across the dataset. This two-step process enhances semantic consistency, transitioning from class-agnostic to more meaningful segmentations, all without relying on labels.
• Figure 1: Additional results for Carotid dataset (Ours$_{proc}$).
• Figure 2: Combining ultrasound-based affinities and deep features leads to meaningful image separation.
• Figure :
• Figure :