Umeå University's logo

Detecting and segmenting arterial plaques in ultrasound images is essential for the early diagnosis and prevention of cardiovascular diseases. This paper presents Hybrid Stable AnoGAN (HS-AnoGAN), an enhanced anomaly detection framework based on AnoGAN (Anomaly Generative Adversarial Network), which utilizes generative adversarial learning to model normal anatomical structures and identify abnormal regions indicative of pathology. The proposed approach introduces key improvements, including direct latent space encoding, hybrid reconstruction loss, feature matching in the discriminator, and adaptive thresholding, leading to more precise anomaly localization. Additionally, spectral normalization and Wasserstein loss with gradient penalty are incorporated to improve training stability and prevent mode collapse. To the best of our knowledge, this is the first attempt to apply anomaly detection techniques for arterial plaque detection and segmentation in ultrasound images. Comparative experiments demonstrate that HS-AnoGAN outperforms state-of-the-art methods, achieving a 9.8% increase in detection accuracy, and a 7.5% enhancement in Dice score for segmentation quality. These results highlight the effectiveness of HS-AnoGAN in improving both plaque detection and segmentation in ultrasound imaging, making it a promising tool for clinical applications.

Background/Introduction: Cardiovascular diseases (CVDs) are the leading cause of death worldwide, with atherosclerosis being a major contributor. Early detection of carotid artery plaques is crucial for CVD prevention and management. While ultrasound imaging is widely used, its manual analysis is time-consuming and prone to variability. Automating plaque detection with deep learning can improve diagnostic consistency and efficiency. Recent advancements in unsupervised anomaly detection models offer a promising approach to detect plaque-related anomalies in carotid ultrasound images without labeled data. This study is, to the best of our knowledge, the first to apply these models for plaque detection in ultrasound images.

Purpose: This work aims to evaluate the effectiveness of three unsupervised anomaly detection methods—Autoencoder (AE), Variational Autoencoder (VAE), and Anomaly Detection Generative Adversarial Networks (AnoGAN)—for detecting plaques in longitudinal carotid artery ultrasound images. We used carotid ultrasound images from a large Randomized Controlled Trial (RCT) on subclinical atherosclerosis, comprising images from a diverse group of participants collected at both baseline and follow-up periods.

Methods: The generative models were trained to recognize normal anatomical structures using 3,000 plaque-free (normal) images. These models were then tested on 400 images, which included both normal and abnormal images, to distinguish them and detect anomalies in abnormal images. As shown in Figure 1, the anomaly scores were computed by comparing the original images to the generated ones, with a distribution-based cutoff applied to classify the images as normal or abnormal. The threshold for this cutoff was determined based on the distribution of the anomaly scores.

Results: AnoGAN achieved the highest accuracy, as shown in Figure 2, which compares the input image and generated output of AnoGAN, VAE, and AE for an abnormal image. Also, AnoGAN achieved the highest Area Under the Curve of the Receiver Operating Characteristic (AUC-ROC) at 91%, followed by VAE at 79% and AE at 66%. Additionally, AnoGAN achieved the best sensitivity-specificity balance, resulting in the highest AUC-ROC, while VAE demonstrated moderate performance, and AE, despite lower sensitivity, maintained competitive specificity.

Conclusion(s): This study shows the effectiveness of anomaly detection models, especially AnoGAN, in automating the detection of plaques in carotid ultrasound imaging. These models have the potential to reduce diagnostic variability, improve early detection of CVD risk factors, and enhance clinical efficiency. Future research should aim to expand the use of these models to larger datasets and explore their clinical integration to assess real-world applicability. Additionally, optimizing the computational performance and interpretability of these models will be crucial for their widespread adoption in routine clinical settings.

Objective: Accurate sleep-stage classification is crucial for advancing both sleep research and healthcare applications. Traditional deep learning (DL) and domain adaptation (DA) methods often struggle due to the limited availability of labeled data in the target domain and their inability to capture the subtle distinctions between sleep-stage classes, which hampers classification accuracy.

Approach: To address these limitations, we introduce a novel framework, adversarial domain adaptation with active deep learning (ADAADL). This framework combines adversarial learning with active learning (AL) strategies to improve feature alignment and effectively leverage unlabeled data. ADAADL employs two separate sleep-stage classifiers as discriminators, allowing for a more refined consideration of class boundaries during the feature alignment process. Moreover, it incorporates entropy measures alongside cross-entropy loss during training to make better use of the information from unlabeled data. The AL component (ADL) further enhances performance by iteratively selecting and labeling the most informative data points, thereby reducing annotation efforts and improving generalization to unseen data.

Main results: Experimental evaluations on three benchmark EEG datasets demonstrate that ADAADL produces robust, transferable features, significantly outperforming existing DA methods in classification accuracy. This research advances sleep-stage classification techniques, offering improved accuracy for real-world applications and contributing to a deeper understanding of sleep dynamics.

Significance: The proposed ADAADL framework advances the state of the art in sleep-stage classification by effectively leveraging unlabeled data and reducing labeling costs. It offers a scalable and accurate solution for real-world sleep monitoring applications and contributes to a deeper understanding of sleep dynamics through improved modeling of sleep stages.

This paper investigates a method addressing the unsupervised segmentation and joint data augmentation in medical ultrasound imaging based on the modified CycleGAN. Accurate quantification of fascia and muscle is the key for the diagnostics of neuromuscular disorders based on the analysis of image sequences of skeletal muscle contraction. Although the Deep Learning (DL) models represent encouraging results, some challenges exist. The traditional models don't consider the complex interaction between tissues within a muscle and its surroundings, which reduces the performance of the fascia segmentation. Also, the DL requires many annotated datasets, which ignores dealing with noisy and complex ultrasound images. To overcome these issues, we propose a method to generate realistic images, and then present an unsupervised fascia segmentation method. The results show that our method improves the segmentation accuracy in noisy and complex ultrasound images compared to the traditional methods.