Ultrasound Image Dataset Research Articles

Enhancing Thyroid Nodule Detection in Ultrasound Images: A Novel YOLOv8 Architecture with a C2fA Module and Optimized Loss Functions

Thyroid-related diseases, particularly thyroid cancer, are rising globally, emphasizing the critical need for the early detection and accurate screening of thyroid nodules. Ultrasound imaging has inherent limitations—high noise, low contrast, and blurred boundaries—that make manual interpretation subjective and error-prone. To address these challenges, YOLO-Thyroid, an improved model for the automatic detection of thyroid nodules in ultrasound images, is presented herein. Building upon the YOLOv8 architecture, YOLO-Thyroid introduces the C2fA module—an extension of C2f that incorporates Coordinate Attention (CA)—to enhance feature extraction. Additionally, loss functions were incorporated, including class-weighted binary cross-entropy to alleviate class imbalance and SCYLLA-IoU (SIoU) to improve localization accuracy during boundary regression. A publicly available thyroid ultrasound image dataset was optimized using format conversion and data augmentation. The experimental results demonstrate that YOLO-Thyroid outperforms mainstream object detection models across multiple metrics, achieving a higher detection precision of 54%. The recall, calculated based on the detection of nodules containing at least one feature suspected of being malignant, reaches 58.2%, while the model maintains a lightweight structure. The proposed method significantly advances ultrasound nodule detection, providing an effective and practical solution for enhancing diagnostic accuracy in medical imaging.

AnkleImage - An ultrafast ultrasound image dataset to understand the ankle joint muscle contractility

The role of the human ankle joint in activities of daily living, including walking, maintaining balance, and participating in sports, is of paramount importance. Ankle joint dorsiflexion and plantarflexion functionalities mainly account for ground clearance and propulsion power generation during locomotion tasks, where those functionalities are driven by the contraction of ankle joint skeleton muscles. Studies of corresponding muscle contractility during ankle dynamic functions will facilitate us to better understand the joint torque/power generation mechanism, better diagnose potential muscular disorders on the ankle joint, or better develop wearable assistive/rehabilitative robotic devices that assist in community ambulation. This data descriptor reports a new dataset that includes the ankle joint kinematics/kinetics, associated muscle surface electromyography, and ultrafast ultrasound images with various annotations, such as pennation angle, fascicle length, tissue displacements, echogenicity, and muscle thickness, of ten healthy participants when performing volitional isometric, isokinetic, and dynamic ankle joint functions (walking at multiple treadmill speeds, including 0.50 m/s, 0.75 m/s, 1.00 m/s, 1.25 m/s, and 1.50 m/s). Data were recorded by a research-use ultrasound machine, a self-designed ankle testbed, an inertia measurement unit system, a Vicon motion capture system, a surface electromyography system, and an instrumented treadmill. The descriptor in this work presents the results of a data curation or collection exercise from previous works, rather than describing a novel primary/experimental data collection.

Performance of automated machine learning in detecting fundus diseases based on ophthalmologic B-scan ultrasound images

AimTo evaluate the efficacy of automated machine learning (AutoML) models in detecting fundus diseases using ocular B-scan ultrasound images.MethodsOphthalmologists annotated two B-scan ultrasound image datasets to develop three AutoML models—single-label,...

Breast Cancer Prediction Using Transfer Learning-Based Classification Model

Breast cancer is currently the most prevalent type of cancer in women, with a growing number of fatalities worldwide. Different imaging methods like mammography, computed tomography, Magnetic Resonance Imaging, ultrasound, and biopsies assist in detecting breast cancer. Recent developments in deep learning have revolutionized breast cancer pathology by facilitating accurate image categorization. This study introduces a novel approach to enhance detection and classification using the Convolutional Neural Network Deep Learning method and Transfer Learning to create a high-speed, accurate image classification model. The model is trained on pre-processed data subjected to thorough analysis and augmentation to ensure the quality of inputs. The experimental results from the Breast Ultrasound Image dataset indicate that our model, with a 0.1 test size ratio, outperforms its counterparts. It achieved an accuracy of 90.12%, with a loss of 0.2641, validation accuracy of 90.15%, and validation loss of 0.31, evidencing its superior classification capability. This research introduces an innovative approach to the automated diagnosis of breast cancer. By combining CNN, Transfer Learning, and data augmentation, we have developed a desktop application that expedites the classification process and significantly improves accuracy. This advancement represents a key development in machine learning applications for breast cancer prognostics and diagnostics. Doi: 10.28991/ESJ-2024-08-06-014 Full Text: PDF

BREAST LESION SEGMENTATION AND CLASSIFICATION USING U-NET SALIENCY ESTIMATION AND EXPLAINABLE RESIDUAL CONVOLUTIONAL NEURAL NETWORK

Breast cancer (BrC) is one of the most common causes of death among women worldwide. Images of the breast (mammography or ultrasound) may show an anomaly that represents early indicators of BrC. However, accurate breast image interpretation necessitates labor-intensive procedures and highly skilled medical professionals. As a second opinion for the physician, deep learning (DL) tools can be useful for the diagnosis and classification of malignant and benign lesions. However, due to the lack of interpretability of DL algorithms, it is not easy to understand by experts as to how to predict a label. In this work, we proposed multitask U-Net Saliency estimation and DL model-based breast lesion segmentation and classification using ultrasound images. A new contrast enhancement technique is proposed to improve the quality of original images. After that, a new technique was proposed called UNET-Saliency map for the segmentation of breast lesions. Simultaneously, a MobileNetV2 deep model is fine-tuned with additional residual blocks and trained from scratch using original and enhanced images. The purpose of additional blocks is to reduce the number of parameters and better learning of ultrasound images. Training is performed from scratch and extracted features from the deeper layers of both models. In the later step, a new cross-entropy controlled sine-cosine algorithm is developed and selected best features. The main purpose of this step is the reduction of irrelevant features for the classification phase. The selected features are fused in the next step by employing a serial-based Manhattan Distance (SbMD) approach and classified the resultant vector using machine learning classifiers. The results indicate that a wide neural network (W-NN) obtained the highest accuracy of 98.9% and sensitivity rate of 98.70% on the selected breast ultrasound image dataset. The comparison of the proposed method accuracy is conducted with state-of-the-art (SoArt) techniques which show the improved performance.

UCapsNet: A Two-Stage Deep Learning Model Using U-Net and Capsule Network for Breast Cancer Segmentation and Classification in Ultrasound Imaging.

Background/Objectives: Breast cancer remains one of the biggest health challenges for women worldwide, and early detection can be truly lifesaving. Although ultrasound imaging is commonly used to detect tumors, the images are not always of sufficient quality, and, thus, traditional U-Net models often miss the finer details needed for accurate detection. This outcome can result in lower accuracy, making early and precise diagnosis more difficult. Methods: This study presents an enhanced U-Net model integrated with a Capsule Network (called UCapsNet) to overcome the limitations of conventional techniques. Our approach improves segmentation by leveraging higher filter counts and skip connections, while the capsule network enhances classification by preserving spatial hierarchies through dynamic routing. The proposed UCapsNet model operates in two stages: first, it segments tumor regions using an enhanced U-Net, followed by a classification of the segmented images with the capsule network. Results: We have tested our model against well-known pre-trained models, including VGG-19, DenseNet, MobileNet, ResNet-50, and Xception. By properly addressing the limitations found in previous studies and using a capsule network trained on the Breast Ultrasound Image (BUSI) dataset, our model resulted in top-achieving impressive precision, recall, and accuracy rates of 98.12%, 99.52%, and 99.22%, respectively. Conclusions: By combining the U-Net's powerful segmentation capabilities with the capsule network's high classification accuracy, UCapsNet boosts diagnostic precision and addresses key weaknesses in existing methods. The findings demonstrate that the proposed model is not only more effective in detecting tumors but also more reliable for practical applications in clinical settings.

Deep Learning Model for Quality Assessment of Urinary Bladder Ultrasound Images using Multi-scale and Higher-order Processing.

Autonomous Ultrasound Image Quality Assessment (US-IQA) is a promising tool to aid the interpretation by practicing sonographers and to enable the future robotization of ultrasound procedures. However, autonomous US-IQA has several challenges. Ultrasound images contain many spurious artifacts, such as noise due to handheld probe positioning, errors in the selection of probe parameters and patient respiration during the procedure. Further, these images are highly variable in appearance with respect to the individual patient's physiology. We propose to use a deep Convolutional Neural Network (CNN), USQNet, which utilizes a Multi-scale and Local-to-Global Second-order Pooling (MS-L2GSoP) classifier to conduct the sonographer-like assessment of image quality. This classifier first extracts features at multiple scales to encode the inter-patient anatomical variations, similar to a sonographer's understanding of anatomy. Then, it uses second-order pooling in the intermediate layers (local) and at the end of the network (global) to exploit the second-order statistical dependency of multi-scale structural and multi-region textural features. The L2GSoP will capture the higher-order relationships between different spatial locations and provide the seed for correlating local patches, much like a sonographer prioritizes regions across the image. We experimentally validated the USQNet for a new dataset of the human urinary bladder ultrasound images. The validation involved first with the subjective assessment by experienced radiologists' annotation, and then with state-of-the-art CNN networks for US-IQA and its ablated counterparts. The results demonstrate that USQNet achieves a remarkable accuracy of 92.4% and outperforms the SOTA models by 3 - 14% while requiring comparable computation time.

Synthesizing Real-Time Ultrasound Images of Muscle Based on Biomechanical Simulation and Conditional Diffusion Network.

Quantitative muscle function analysis based on the ultrasound imaging, has been used for various applications, particularly with recent development of deep learning methods. The nature of speckle noises in ultrasound images poses challenges to accurate and reliable data annotation for supervised learning algorithms. To obtain a large and reliable dataset without manual scanning and labelling, we proposed a synthesizing pipeline to provide synthetic ultrasound datasets of muscle movement with an accurate ground truth, allowing augmenting, training, and evaluating models for different tasks. Our pipeline contained biomechanical simulation using finite element method, an algorithm for reconstructing sparse fascicles, and a diffusion network for ultrasound image generation. With the adjustment of a few parameters, the proposed pipeline can generate a large dataset of real-time ultrasound images with diversity in morphology and pattern. With 3,030 ultrasound images generated, we qualitatively and quantitatively verified that the synthetic images closely matched with the in-vivo images. In addition, we applied the synthetic dataset into different tasks of muscle analysis. Compared to trained on an unaugmented dataset, model trained on synthetic one had better cross-dataset performance, which demonstrates the feasibility of synthesizing pipeline to augment model training and avoid over-fitting. The results of the regression task show potentials under the conditions that the number of datasets or the accurate label are limited. The proposed synthesizing pipeline can not only be used for muscle-related study, but for other similar study and model development, where sequential images are needed for training.

Achieving enhanced diagnostic precision in endometrial lesion analysis through a data enhancement framework.

The aim of this study was to enhance the precision of categorization of endometrial lesions in ultrasound images via a data enhancement framework based on deep learning (DL), through addressing diagnostic accuracy challenges, contributing to future research. Ultrasound image datasets from 734 patients across six hospitals were collected. A data enhancement framework, including image features cleaning and soften label, was devised and validated across multiple DL models, including ResNet50, DenseNet169, DenseNet201, and ViT-B. A hybrid model, integrating convolutional neural network and transformer architectures for optimal performance, to predict lesion types was developed. Implementation of our novel strategies resulted in a substantial enhancement in model accuracy. The ensemble model achieved accuracy and macro-area under the receiver operating characteristic curve values of 0.809 of 0.911, respectively, underscoring the potential for use of DL in endometrial lesion ultrasound image classification. We successfully developed a data enhancement framework to accurately classify endometrial lesions in ultrasound images. Integration of anomaly detection, data cleaning, and soften label strategies enhanced the comprehension of lesion image features by the model, thereby boosting its classification capacity. Our research offers valuable insights for future studies and lays the foundation for creation of more precise diagnostic tools.

Deep learning-driven ultrasound-assisted diagnosis: optimizing GallScopeNet for precise identification of biliary atresia.

Biliary atresia (BA) is a severe congenital biliary developmental abnormality threatening neonatal health. Traditional diagnostic methods rely heavily on experienced radiologists, making the process time-consuming and prone to variability. The application of deep learning for the automated diagnosis of BA remains underexplored. This study introduces GallScopeNet, a deep learning model designed to improve diagnostic efficiency and accuracy through innovative architecture and advanced feature extraction techniques. The model utilizes data from a carefully constructed dataset of gallbladder ultrasound images. A dataset comprising thousands of ultrasound images was employed, with the majority used for training and validation and a subset reserved for external testing. The model's performance was evaluated using five-fold cross-validation and external assessment, employing metrics such as accuracy and the area under the receiver operating characteristic curve (AUC), compared against clinical diagnostic standards. GallScopeNet demonstrated exceptional performance in distinguishing BA from non-BA cases. In the external test dataset, GallScopeNet achieved an accuracy of 81.21% and an AUC of 0.85, indicating strong diagnostic capabilities. The results highlighted the model's ability to maintain high classification performance, reducing misdiagnosis and missed diagnosis. GallScopeNet effectively differentiates between BA and non-BA images, demonstrating significant potential and reliability for early diagnosis. The system's high efficiency and accuracy suggest it could serve as a valuable diagnostic tool in clinical settings, providing substantial technical support for improving diagnostic workflows.

Attention based UNet model for breast cancer segmentation using BUSI dataset.

Breast cancer, a prevalent and life-threatening disease, necessitates early detection for the effective intervention and the improved patient health outcomes. This paper focuses on the critical problem of identifying breast cancer using a model called Attention U-Net. The model is utilized on the Breast Ultrasound Image Dataset (BUSI), comprising 780 breast images. The images are categorized into three distinct groups: 437 cases classified as benign, 210 cases classified as malignant, and 133 cases classified as normal. The proposed model leverages the attention-driven U-Net's encoder blocks to capture hierarchical features effectively. The model comprises four decoder blocks which is a pivotal component in the U-Net architecture, responsible for expanding the encoded feature representation obtained from the encoder block and for reconstructing spatial information. Four attention gates are incorporated strategically to enhance feature localization during decoding, showcasing a sophisticated design that facilitates accurate segmentation of breast tumors in ultrasound images. It displays its efficacy in accurately delineating and segregating tumor borders. The experimental findings demonstrate outstanding performance, achieving an overall accuracy of 0.98, precision of 0.97, recall of 0.90, and a dice score of 0.92. It demonstrates its effectiveness in precisely defining and separating tumor boundaries. This research aims to make automated breast cancer segmentation algorithms by emphasizing the importance of early detection in boosting diagnostic capabilities and enabling prompt and targeted medical interventions.

Incorporation of Structural Similarity Index and Regularization Term into Neighbor2Neighbor Unsupervised Learning Model for Efficient Ultrasound Image Data Denoising

Medical ultrasound imaging is extensively employed for diagnostic purposes. However, image quality remains a major obstacle to achieving greater accuracy. Conventional supervised deep learning denoising methods often rely on matched noise-free and noisy image pairs, which can be highly challenging in practical ultrasound applications. Moreover, due to the limitations associated with independent noise, existing unsupervised denoising methods such as Neighbor2Neighbor are unable to efficiently address correlated noise in ultrasound images. Meanwhile, these methods utilize a random neighborhood downsampling technique, frequently resulting in pixel loss. Hence, this study proposes a novel Neighbor2Neighbor algorithm, which reconstructs ultrasound images by improving the downsampling approach. Moreover, it incorporates a structural similarity index and a regularization term, thereby enhancing its ability to suppress both independent and correlated noise. Extensive experiments on an ultrasound image dataset demonstrate that the proposed Neighbor2Neighbor algorithm outperforms the state-of-the-art baseline algorithms in peak signal-to-noise ratio (PSNR), mean structural similarity index measure (MSSIM), feature similarity index measure (FSIM), and edge preservation index (EPI).

Automatic Classification of Nodules from 2D Ultrasound Images Using Deep Learning Networks.

In clinical practice, thyroid nodules are typically visually evaluated by expert physicians using 2D ultrasound images. Based on their assessment, a fine needle aspiration (FNA) may be recommended. However, visually classifying thyroid nodules from ultrasound images may lead to unnecessary fine needle aspirations for patients. The aim of this study is to develop an automatic thyroid ultrasound image classification system to prevent unnecessary FNAs. An automatic computer-aided artificial intelligence system is proposed for classifying thyroid nodules using a fine-tuned deep learning model based on the DenseNet architecture, which incorporates an attention module. The dataset comprises 591 thyroid nodule images categorized based on the Bethesda score. Thyroid nodules are classified as either requiring FNA or not. The challenges encountered in this task include managing variability in image quality, addressing the presence of artifacts in ultrasound image datasets, tackling class imbalance, and ensuring model interpretability. We employed techniques such as data augmentation, class weighting, and gradient-weighted class activation maps (Grad-CAM) to enhance model performance and provide insights into decision making. Our approach achieved excellent results with an average accuracy of 0.94, F1-score of 0.93, and sensitivity of 0.96. The use of Grad-CAM gives insights on the decision making and then reinforce the reliability of the binary classification for the end-user perspective. We propose a deep learning architecture that effectively classifies thyroid nodules as requiring FNA or not from ultrasound images. Despite challenges related to image variability, class imbalance, and interpretability, our method demonstrated a high classification accuracy with minimal false negatives, showing its potential to reduce unnecessary FNAs in clinical settings.

DFA-UNet: dual-stream feature-fusion attention U-Net for lymph node segmentation in lung cancer diagnosis.

In bronchial ultrasound elastography, accurately segmenting mediastinal lymph nodes is of great significance for diagnosing whether lung cancer has metastasized. However, due to the ill-defined margin of ultrasound images and the complexity of lymph node structure, accurate segmentation of fine contours is still challenging. Therefore, we propose a dual-stream feature-fusion attention U-Net (DFA-UNet). Firstly, a dual-stream encoder (DSE) is designed by combining ConvNext with a lightweight vision transformer (ViT) to extract the local information and global information of images; Secondly, we propose a hybrid attention module (HAM) at the bottleneck, which incorporates spatial and channel attention to optimize the features transmission process by optimizing high-dimensional features at the bottom of the network. Finally, the feature-enhanced residual decoder (FRD) is developed to improve the fusion of features obtained from the encoder and decoder, ensuring a more comprehensive integration. Extensive experiments on the ultrasound elasticity image dataset show the superiority of our DFA-UNet over 9 state-of-the-art image segmentation models. Additionally, visual analysis, ablation studies, and generalization assessments highlight the significant enhancement effects of DFA-UNet. Comprehensive experiments confirm the excellent segmentation effectiveness of the DFA-UNet combined attention mechanism for ultrasound images, underscoring its important significance for future research on medical images.

HCN: Hybrid Capsule Network for Fetal Plane Classification in Ultrasound Images

ABSTRACTClassifying fetal ultrasound images into different anatomical categories, such as the abdomen, brain, femur, thorax, and so forth can contribute to the early identification of potential anomalies or dangers during prenatal care. Ignoring major abnormalities that might lead to fetal death or permanent disability. This article proposes a novel hybrid capsule network architecture‐based method for identifying fetal ultrasound images. The proposed architecture increases the precision of fetal image categorization by combining the benefits of a capsule network with a convolutional neural network. The proposed hybrid model surpasses conventional convolutional network‐based techniques with an overall accuracy of 0.989 when tested on a publicly accessible dataset of prenatal ultrasound images. The results indicate that the proposed hybrid architecture is a promising approach for precisely and consistently classifying fetal ultrasound images, with potential uses in clinical settings.

A deep image classification model based on prior feature knowledge embedding and application in medical diagnosis

Aiming at the problem of image classification with insignificant morphological structural features, strong target correlation, and low signal-to-noise ratio, combined with prior feature knowledge embedding, a deep learning method based on ResNet and Radial Basis Probabilistic Neural Network (RBPNN) is proposed model. Taking ResNet50 as a visual modeling network, it uses feature pyramid and self-attention mechanism to extract appearance and semantic features of images at multiple scales, and associate and enhance local and global features. Taking into account the diversity of category features, channel cosine similarity attention and dynamic C-means clustering algorithms are used to select representative sample features in different category of sample subsets to implicitly express prior category feature knowledge, and use them as the kernel centers of radial basis probability neurons (RBPN) to realize the embedding of diverse prior feature knowledge. In the RBPNN pattern aggregation layer, the outputs of RBPN are selectively summed according to the category of the kernel center, that is, the subcategory features are combined into category features, and finally the image classification is implemented based on Softmax. The functional module of the proposed method is designed specifically for image characteristics, which can highlight the significance of local and structural features of the image, form a non-convex decision-making area, and reduce the requirements for the completeness of the sample set. Applying the proposed method to medical image classification, experiments were conducted based on the brain tumor MRI image classification public dataset and the actual cardiac ultrasound image dataset, and the accuracy rate reached 85.82% and 83.92% respectively. Compared with the three mainstream image classification models, the performance indicators of this method have been significantly improved.

Integration of feature enhancement technique in Google inception network for breast cancer detection and classification

Breast cancer is a major public health concern, and early detection and classification are essential for improving patient outcomes. However, breast tumors can be difficult to distinguish from benign tumors, leading to high false positive rates in screening. The reason is that both benign and malignant tumors have no consistent shape, are found at the same position, have variable sizes, and have high correlations. The ambiguity of the correlation challenges the computer-aided system, and the inconsistency of morphology challenges an expert in identifying and classifying what is positive and what is negative. Due to this, most of the time, breast cancer screen is prone to false positive rates. This research paper presents the introduction of a feature enhancement method into the Google inception network for breast cancer detection and classification. The proposed model preserves both local and global information, which is important for addressing the variability of breast tumor morphology and their complex correlations. A locally preserving projection transformation function is introduced to retain local information that might be lost in the intermediate output of the inception model. Additionally, transfer learning is used to improve the performance of the proposed model on limited datasets. The proposed model is evaluated on a dataset of ultrasound images and achieves an accuracy of 99.81%, recall of 96.48%, and sensitivity of 93.0%. These results demonstrate the effectiveness of the proposed method for breast cancer detection and classification.

PSFHS: Intrapartum ultrasound image dataset for AI-based segmentation of pubic symphysis and fetal head

During the process of labor, the intrapartum transperineal ultrasound examination serves as a valuable tool, allowing direct observation of the relative positional relationship between the pubic symphysis and fetal head (PSFH). Accurate assessment of fetal head descent and the prediction of the most suitable mode of delivery heavily rely on this relationship. However, achieving an objective and quantitative interpretation of the ultrasound images necessitates precise PSFH segmentation (PSFHS), a task that is both time-consuming and demanding. Integrating the potential of artificial intelligence (AI) in the field of medical ultrasound image segmentation, the development and evaluation of AI-based models rely significantly on access to comprehensive and meticulously annotated datasets. Unfortunately, publicly accessible datasets tailored for PSFHS are notably scarce. Bridging this critical gap, we introduce a PSFHS dataset comprising 1358 images, meticulously annotated at the pixel level. The annotation process adhered to standardized protocols and involved collaboration among medical experts. Remarkably, this dataset stands as the most expansive and comprehensive resource for PSFHS to date.

RTSeg-net: A lightweight network for real-time segmentation of fetal head and pubic symphysis from intrapartum ultrasound images

The segmentation of the fetal head (FH) and pubic symphysis (PS) from intrapartum ultrasound images plays a pivotal role in monitoring labor progression and informing crucial clinical decisions. Achieving real-time segmentation with high accuracy on systems with limited hardware capabilities presents significant challenges. To address these challenges, we propose the real-time segmentation network (RTSeg-Net), a groundbreaking lightweight deep learning model that incorporates innovative distribution shifting convolutional blocks, tokenized multilayer perceptron blocks, and efficient feature fusion blocks. Designed for optimal computational efficiency, RTSeg-Net minimizes resource demand while significantly enhancing segmentation performance. Our comprehensive evaluation on two distinct intrapartum ultrasound image datasets reveals that RTSeg-Net achieves segmentation accuracy on par with more complex state-of-the-art networks, utilizing merely 1.86 M parameters—just 6 % of their hyperparameters—and operating seven times faster, achieving a remarkable rate of 31.13 frames per second on a Jetson Nano, a device known for its limited computing capacity. These achievements underscore RTSeg-Net's potential to provide accurate, real-time segmentation on low-power devices, broadening the scope for its application across various stages of labor. By facilitating real-time, accurate ultrasound image analysis on portable, low-cost devices, RTSeg-Net promises to revolutionize intrapartum monitoring, making sophisticated diagnostic tools accessible to a wider range of healthcare settings.

On the use of contrastive learning for standard-plane classification in fetal ultrasound imaging

Background:To investigate the effectiveness of contrastive learning, in particular SimClr, in reducing the need for large annotated ultrasound (US) image datasets for fetal standard plane identification. Methods:We explore SimClr advantage in the cases of both low and high inter-class variability, considering at the same time how classification performance varies according to different amounts of labels used. This evaluation is performed by exploiting contrastive learning through different training strategies. We apply both quantitative and qualitative analyses, using standard metrics (F1-score, sensitivity, and precision), Class Activation Mapping (CAM), and t-Distributed Stochastic Neighbor Embedding (t-SNE). Results:When dealing with high inter-class variability classification tasks, contrastive learning does not bring a significant advantage; whereas it results to be relevant for low inter-class variability classification, specifically when initialized with ImageNet weights. Conclusions:Contrastive learning approaches are typically used when a large number of unlabeled data is available, which is not representative of US datasets. We proved that SimClr either as pre-training with backbone initialized via ImageNet weights or used in an end-to-end dual-task may impact positively the performance over standard transfer learning approaches, under a scenario in which the dataset is small and characterized by low inter-class variability.