Medical ultrasonography is an essential imaging technique for examining superficial organs and tissues, including lymph nodes, breast, and thyroid. It employs high-frequency ultrasound waves to generate detailed images of the internal structures of the human body. However, manually contouring regions of interest in these images is a labor-intensive task that demands expertise and often results in inconsistent interpretations among individuals. Vision-language foundation models, which have excelled in various computer vision applications, present new opportunities for enhancing ultrasound image analysis. Yet, their performance is hindered by the significant differences between natural and medical imaging domains. This research seeks to overcome these challenges by developing domain adaptation methods for vision-language foundation models. In this study, we explore the fine-tuning pipeline for vision-language foundation models by utilizing large language model as text refiner with special-designed adaptation strategies and task-driven heads. Our approach has been extensively evaluated on six ultrasound datasets and two tasks: segmentation and classification. The experimental results show that our method can effectively improve the performance of vision-language foundation models for ultrasound image analysis, and outperform the existing state-of-the-art vision-language and pure foundation models. The source code of this study is available at https://github.com/jinggqu/NextGen-UIA.

Accurate classification of second-trimester fetal ultrasound images remains challenging due to low image quality, high intra-class variability, and significant class imbalance. In this work, we introduce a simple yet powerful, biologically inspired deep learning ensemble framework that-unlike prior studies focused on only a handful of anatomical targets-simultaneously distinguishes 16 fetal structures. Drawing on the hierarchical, modular organization of biological vision systems, our model stacks two complementary branches (a "shallow" path for coarse, low-resolution cues and a "detailed" path for fine, high-resolution features), concatenating their outputs for final prediction. To our knowledge, no existing method has addressed such a large number of classes with a comparably lightweight architecture. We trained and evaluated on 5,298 routinely acquired clinical images (annotated by three experts and reconciled via Dawid-Skene), reflecting real-world noise and variability rather than a "cleaned" dataset. Despite this complexity, our ensemble (EfficientNet-B0 + EfficientNet-B6 with LDAM-Focal loss) identifies 90% of organs with accuracy > 0.75 and 75% of organs with accuracy > 0.85-performance competitive with more elaborate models applied to far fewer categories. These results demonstrate that biologically inspired modular stacking can yield robust, scalable fetal anatomy recognition in challenging clinical settings.

This study aimed to differentiate between benign and malignant gallbladder polyps preoperatively by developing a prediction model integrating preoperative transabdominal ultrasound and clinical features using machine-learning algorithms. A retrospective analysis was conducted on clinical and ultrasound data from 1,050 patients at 2 centers who underwent cholecystectomy for gallbladder polyps. Six machine-learning algorithms were used to develop preoperative models for predicting benign and malignant gallbladder polyps. Internal and external test cohorts evaluated model performance. The Shapley Additive Explanations algorithm was used to understand feature importance. The main study cohort included 660 patients with benign polyps and 285 patients with malignant polyps, randomly divided into a 3:1 stratified training and internal test cohorts. The external test cohorts consisted of 73 benign and 32 malignant polyps. In the training cohort, the Shapley Additive Explanations algorithm, on the basis of variables selected by Least Absolute Shrinkage and Selection Operator regression and multivariate logistic regression, further identified 6 key predictive factors: polyp size, age, fibrinogen, carbohydrate antigen 19-9, presence of stones, and cholinesterase. Using these factors, 6 predictive models were developed. The random forest model outperformed others, with an area under the curve of 0.963, 0.940, and 0.958 in the training, internal, and external test cohorts, respectively. Compared with previous studies, the random forest model demonstrated excellent clinical utility and predictive performance. In addition, the Shapley Additive Explanations algorithm was used to visualize feature importance, and an online calculation platform was developed. The random forest model, combining preoperative ultrasound and clinical features, accurately predicts benign and malignant gallbladder polyps, offering valuable guidance for clinical decision-making.

Anal injuries, such as lacerations and fissures, are challenging to diagnose because of their anatomical complexity. Endoanal ultrasound (EAUS) has proven to be a reliable tool for detailed visualization of anal structures but relies on expert interpretation. Artificial intelligence (AI) may offer a solution for more accurate and consistent diagnoses. This study aims to develop and test a convolutional neural network (CNN)-based algorithm for automatic classification of fissures and anal lacerations (internal and external) on EUAS. A single-center retrospective study analyzed 238 EUAS radial probe exams (April 2022-January 2024), categorizing 4528 frames into fissures (516), external lacerations (2174), and internal lacerations (1838), following validation by three experts. Data was split 80% for training and 20% for testing. Performance metrics included sensitivity, specificity, and accuracy. For external lacerations, the CNN achieved 82.5% sensitivity, 93.5% specificity, and 88.2% accuracy. For internal lacerations, achieved 91.7% sensitivity, 85.9% specificity, and 88.2% accuracy. For anal fissures, achieved 100% sensitivity, specificity, and accuracy. This first EUAS AI-assisted model for differentiating benign anal injuries demonstrates excellent diagnostic performance. It highlights AI's potential to improve accuracy, reduce reliance on expertise, and support broader clinical adoption. While currently limited by small dataset and single-center scope, this work represents a significant step towards integrating AI in proctology.

Ultrasound (US) imaging is a promising modality for real-time monitoring of robotic capsule endoscopes navigating through the gastrointestinal (GI) tract. It offers high temporal resolution and safety but is limited by a narrow field of view, low visibility in gas-filled regions and challenges in detecting out-of-plane motions. This work addresses these issues by proposing a novel robotic ultrasound tracking system capable of long-distance 3D tracking and active re-localization when the capsule is lost due to motion or artifacts. We develop a hybrid deep learning-based tracking framework combining convolutional neural networks (CNNs) and a transformer backbone. The CNN component efficiently encodes spatial features, while the transformer captures long-range contextual dependencies in B-mode US images. This model is integrated with a robotic arm that adaptively scans and tracks the capsule. The system's performance is evaluated using ex vivo colon phantoms under varying imaging conditions, with physical perturbations introduced to simulate realistic clinical scenarios. The proposed system achieved continuous 3D tracking over distances exceeding 90 cm, with a mean centroid localization error of 1.5 mm and over 90% detection accuracy. We demonstrated 3D tracking in a more complex workspace featuring two curved sections to simulate anatomical challenges. This suggests the strong resilience of the tracking system to motion-induced artifacts and geometric variability. The system maintained real-time tracking at 9-12 FPS and successfully re-localized the capsule within seconds after tracking loss, even under gas artifacts and acoustic shadowing. This study presents a hybrid CNN-transformer system for automatic, real-time 3D ultrasound tracking of capsule robots over long distances. The method reliably handles occlusions, view loss and image artifacts, offering millimeter-level tracking accuracy. It significantly reduces clinical workload through autonomous detection and re-localization. Future work includes improving probe-tissue interaction handling and validating performance in live animal and human trials to assess physiological impacts.

Multimodal large language models (MLLMs) have shown great potential in general domains but perform poorly in some specific domains due to a lack of domain-specific data, such as image-text data or vedio-text data. In some specific domains, there is abundant graphic and textual data scattered around, but lacks standardized arrangement. In the field of medical ultrasound, there are ultrasonic diagnostic books, ultrasonic clinical guidelines, ultrasonic diagnostic reports, and so on. However, these ultrasonic materials are often saved in the forms of PDF, images, etc., and cannot be directly used for the training of MLLMs. This paper proposes a novel image-text reasoning supervised fine-tuning data generation pipeline to create specific domain quadruplets (image, question, thinking trace, and answer) from domain-specific materials. A medical ultrasound domain dataset ReMUD is established, containing over 45,000 reasoning and non-reasoning supervised fine-tuning Question Answering (QA) and Visual Question Answering (VQA) data. The ReMUD-7B model, fine-tuned on Qwen2.5-VL-7B-Instruct, outperforms general-domain MLLMs in medical ultrasound field. To facilitate research, the ReMUD dataset, data generation codebase, and ReMUD-7B parameters will be released at https://github.com/ShiDaizi/ReMUD, addressing the data shortage issue in specific domain MLLMs.

The present study aims to construct a random forest (RF) model based on ultrasound radiomics and elastography, offering a new approach for the differentiation of thyroid nodules (TNs). We retrospectively analyzed 152 TNs from 127 patients and developed four machine learning models. The examination was performed using the Resona 9Pro equipped with a 15-4 MHz linear array probe. The region of interest (ROI) was delineated with 3D Slicer. Using the RF algorithm, four models were developed based on sound touch elastography (STE) parameters, strain elastography (SE) parameters, and the selected radiomic features: the STE model, SE model, radiomics model, and the combined model. Decision Curve Analysis (DCA) is employed to assess the clinical benefit of each model. The DeLong test is used to determine whether the area under the curves (AUC) values of different models are statistically significant. A total of 1396 radiomic features were extracted using the Pyradiomics package. After screening, a total of 7 radiomic features were ultimately included in the construction of the model. In STE, SE, radiomics model, and combined model, the AUCs are 0.699 (95% CI: 0.570-0.828), 0.812 (95% CI: 0.683-0.941), 0.851 (95% CI: 0.739-0.964) and 0.911 (95% CI: 0.806-1.000), respectively. In these models, the combined model and the radiomics model exhibited outstanding performance. The combined model, integrating elastography and radiomics, demonstrates superior predictive accuracy compared to single models, offering a promising approach for the diagnosis of TNs.

The purpose of this study is to evaluate the diagnostic accuracy of an artificial intelligence (AI)-based Computer-Aided Diagnosis (CADx) system for breast ultrasound, compare its performance with radiologists, and assess the effect of AI-assisted diagnosis. This study aims to investigate the system's ability to differentiate between benign and malignant breast masses among Japanese patients. This retrospective study included 171 breast mass ultrasound images (92 benign, 79 malignant). The AI system, BU-CAD™, provided Breast Imaging Reporting and Data System (BI-RADS) categorization, which was compared with the performance of three radiologists. Diagnostic accuracy, sensitivity, specificity, and area under the curve (AUC) were analyzed. Radiologists' diagnostic performance with and without AI assistance was also compared, and their reading time was measured using a stopwatch. The AI system demonstrated a sensitivity of 91.1%, specificity of 92.4%, and an AUC of 0.948. It showed comparable diagnostic performance to Radiologist 1, with 10 years of experience in breast imaging (0.948 vs. 0.950; p = 0.893), and superior performance to Radiologist 2 (7 years of experience, 0.948 vs. 0.881; p = 0.015) and Radiologist 3 (3 years of experience, 0.948 vs. 0.832; p = 0.001). When comparing diagnostic performance with and without AI, the use of AI significantly improved the AUC for Radiologists 2 and 3 (p = 0.001 and 0.005, respectively). However, there was no significant difference for Radiologist 1 (p = 0.139). In terms of diagnosis time, the use of AI reduced the reading time for all radiologists. Although there was no significant difference in diagnostic performance between AI and Radiologist 1, the use of AI substantially decreased the diagnosis time for Radiologist 1 as well. The AI system significantly improved diagnostic efficiency and accuracy, particularly for junior radiologists, highlighting its potential clinical utility in breast ultrasound diagnostics.

Stroke, a leading global cause of mortality and neurological disability, is often associated with atherosclerotic carotid artery disease. Distinguishing between symptomatic and asymptomatic carotid artery disease is crucial for appropriate treatment decisions. Radiomics, a quantitative image analysis technique, and machine learning (ML) have emerged as promising tools in Ultrasound (US) imaging, potentially providing a helpful tool in the screening of such lesions. Pubmed, Web of Science and Scopus databases were searched for relevant studies published from January 2005 to May 2023. The Radiomics Quality Score (RQS) was used to assess methodological quality of studies included in the review. The Quality Assessment of Diagnostic Accuracy Studies (QUADAS-2) assessed the risk of bias. Sensitivity, specificity, and logarithmic diagnostic odds ratio (logDOR) meta-analyses have been conducted, alongside an influence analysis. RQS assessed methodological quality, revealing an overall low score and consistent findings with other radiology domains. QUADAS-2 indicated an overall low risk, except for two studies with high bias. The meta-analysis demonstrated that radiomics-based ML models for predicting culprit plaques on US had a satisfactory performance, with a sensitivity of 0.84 and specificity of 0.82. The logDOR analysis confirmed the positive results, yielding a pooled logDOR of 3.54. The summary ROC curve provided an AUC of 0.887. Radiomics combined with ML provide high sensitivity and low false positive rate for carotid plaque vulnerability assessment on US. However, current evidence is not definitive, given the low overall study quality and high inter-study heterogeneity. High quality, prospective studies are needed to confirm the potential of these promising techniques.

Breast cancer is the second deadliest cancer among women after lung cancer. Though the breast cancer death rate continues to decline in the past 20 years, the stages IV and III breast cancer death rates remain high. Therefore, an automated breast cancer diagnosis system is of great significance for early screening of breast lesions to improve the survival rate of patients. This paper proposes a deep learning-based network hybrid adaptive attention deep supervision-guided U-Net (HAA-DSUNet) for breast lesion segmentation of breast ultrasound computed tomography (BUCT) images, which replaces the traditionally sampled convolution module of U-Net with the hybrid adaptive attention module (HAAM), aiming to enlarge the receptive field and probe rich global features while preserving fine details. Moreover, we apply the contrast loss to intermediate outputs as deep supervision to minimize the information loss during upsampling. Finally, the segmentation prediction results are further processed by filtering, segmentation, and morphology to obtain the final results. We conducted the experiment on our two UCT image datasets HCH and HCH-PHMC, and the highest Dice score is 0.8729 and IoU is 0.8097, which outperform all the other state-of-the-art methods. It is demonstrated that our algorithm is effective in segmenting the legion from BUCT images.