Ultrasound Image Dataset Research Articles

Real-Time Deep Recognition of Standardized Liver Ultrasound Scan Locations.

Liver ultrasound (US) plays a critical role in diagnosing liver diseases. However, it is often difficult for examiners to accurately identify the liver segments captured in US images due to patient variability and the complexity of US images. Our study aim is automatic, real-time recognition of standardized US scans coordinated with reference liver segments to guide examiners. We propose a novel deep hierarchical architecture for classifying liver US images into 11 standardized US scans, which has yet to be properly established due to excessive variability and complexity. We address this problem based on a hierarchical classification of 11 US scans with different features applied to individual hierarchies as well as a novel feature space proximity analysis for handling ambiguous US images. Experiments were performed using US image datasets obtained from a hospital setting. To evaluate the performance under patient variability, we separated the training and testing datasets into distinct patient groups. The experimental results show that the proposed method achieved an F1-score of more than 93%, which is more than sufficient for a tool to guide examiners. The superior performance of the proposed hierarchical architecture was demonstrated by comparing its performance with that of non-hierarchical architecture.

Deep learning for classification of thyroid nodules on ultrasound: validation on an independent dataset

ObjectivesThe purpose is to apply a previously validated deep learning algorithm to a new thyroid nodule ultrasound image dataset and compare its performances with radiologists. MethodsPrior study presented an algorithm which is able to detect thyroid nodules and then make malignancy classifications with two ultrasound images. A multi-task deep convolutional neural network was trained from 1278 nodules and originally tested with 99 separate nodules. The results were comparable with that of radiologists. The algorithm was further tested with 378 nodules imaged with ultrasound machines from different manufacturers and product types than the training cases. Four experienced radiologists were requested to evaluate the nodules for comparison with deep learning. ResultsThe Area Under Curve (AUC) of the deep learning algorithm and four radiologists were calculated with parametric, binormal estimation. For the deep learning algorithm, the AUC was 0.69 (95% CI: 0.64–0.75). The AUC of radiologists were 0.63 (95% CI: 0.59–0.67), 0.66 (95% CI:0.61–0.71), 0.65 (95% CI: 0.60–0.70), and 0.63 (95%CI: 0.58–0.67). ConclusionIn the new testing dataset, the deep learning algorithm achieved similar performances with all four radiologists. The relative performance difference between the algorithm and the radiologists is not significantly affected by the difference of ultrasound scanner.

Attention guided neural ODE network for breast tumor segmentation in medical images

Breast cancer is the most common cancer in women. Ultrasound is a widely used screening tool for its portability and easy operation, and DCE-MRI can highlight the lesions more clearly and reveal the characteristics of tumors. They are both noninvasive and nonradiative for assessment of breast cancer. Doctors make diagnoses and further instructions through the sizes, shapes and textures of the breast masses showed on medical images, so automatic tumor segmentation via deep neural networks can to some extent assist doctors. Compared to some challenges which the popular deep neural networks have faced, such as large amounts of parameters, lack of interpretability, overfitting problem, etc., we propose a segmentation network named Att-U-Node which uses attention modules to guide a neural ODE-based framework, trying to alleviate the problems mentioned above. Specifically, the network uses ODE blocks to make up an encoder-decoder structure, feature modeling by neural ODE is completed at each level. Besides, we propose to use an attention module to calculate the coefficient and generate a much refined attention feature for skip connection. Three public available breast ultrasound image datasets (i.e. BUSI, BUS and OASBUD) and a private breast DCE-MRI dataset are used to assess the efficiency of the proposed model, besides, we upgrade the model to 3D for tumor segmentation with the data selected from Public QIN Breast DCE-MRI. The experiments show that the proposed model achieves competitive results compared with the related methods while mitigates the common problems of deep neural networks.

BiTNet: Hybrid deep convolutional model for ultrasound image analysis of human biliary tract and its applications

Certain life-threatening abnormalities, such as cholangiocarcinoma, in the human biliary tract are curable if detected at an early stage, and ultrasonography has been proven to be an effective tool for identifying them. However, the diagnosis often requires a second opinion from experienced radiologists, who are usually overwhelmed by many cases. Therefore, we propose a deep convolutional neural network model, named biliary tract network (BiTNet), developed to solve problems in the current screening system and to avoid overconfidence issues of traditional deep convolutional neural networks. Additionally, we present an ultrasound image dataset for the human biliary tract and demonstrate two artificial intelligence (AI) applications: auto-prescreening and assisting tools. The proposed model is the first AI model to automatically screen and diagnose upper-abdominal abnormalities from ultrasound images in real-world healthcare scenarios. Our experiments suggest that prediction probability has an impact on both applications, and our modifications to EfficientNet solve the overconfidence problem, thereby improving the performance of both applications and of healthcare professionals. The proposed BiTNet can reduce the workload of radiologists by 35% while keeping the false negatives to as low as 1 out of every 455 images. Our experiments involving 11 healthcare professionals with four different levels of experience reveal that BiTNet improves the diagnostic performance of participants of all levels. The mean accuracy and precision of the participants with BiTNet as an assisting tool (0.74 and 0.61, respectively) are statistically higher than those of participants without the assisting tool (0.50 and 0.46, respectively (p<0.001)). These experimental results demonstrate the high potential of BiTNet for use in clinical settings.

DCA-Net: Dual-branch contextual-aware network for auxiliary localization and segmentation of parathyroid glands

Hypoparathyroidism is one of the major complications after thyroid cancer surgery, which seriously affects patients’ quality of life. How to effectively protect parathyroid function and reduce the incidence of hypoparathyroidism has been a key research area in thyroid surgery. Before the operation, the identification and localization of the parathyroid glands by ultrasound images can help surgeons effectively protect parathyroid glands during thyroidectomy. Computer-aided diagnosis (CAD) systems are often considered useful tools to aid in medical image analysis examinations, which provide new solutions for preoperative identification and localization of the parathyroid glands in ultrasound images with high accuracy. To reduce the incidence of hypoparathyroidism and capture both local and global feature dependencies, we propose a dual-branch contextual-aware network (DCA-Net) with Transformer. Firstly, the Transformer as the long-range feature extractor branch is used as the input image patches as the input sequence for extracting the global context. Meanwhile, we build the residual bottleneck block and Transformer bottleneck on the feature encoder branch to locally aggregate the sampled features on different semantic scales. Finally, we use the channel and spatial fusion module (CSFM) to aggregate information from these two branches. The proposed DCA-Net can effectively mitigate the loss of details due to information recession and establish global and local feature dependencies by utilizing these complementing components. Experiments using our dataset of Ultrasound images of Parathyroid Glands reveal that our method outperforms state-of-the-art baselines consistently.

BCDNet: An Optimized Deep Network for Ultrasound Breast Cancer Detection

ObjectivesBreast cancer is a common but deadly disease among women. Medical imaging is an effective method to diagnose breast cancer, but manual image screening is time-consuming. In this study, a novel computer-aided diagnosis system for breast cancer detection called BCDNet is proposed. Material and MethodsWe leverage pre-trained convolutional neural networks (CNNs) for representation learning and propose an adaptive backbone selection algorithm to obtain the best CNN model. An extreme learning machine serves as the classifier in the BCDNet, and a bat algorithm with chaotic maps is put forward to further optimize the parameters in the classifiers. A public ultrasound image dataset is used in the experiments based on 5-fold cross-validation. ResultsSimulation results suggest that our BCDNet outperforms several state-of-the-art breast cancer detection methods in terms of accuracy. ConclusionThe proposed BCDNet is a useful auxiliary tool that can be applied in clinical screening for breast cancer.

Ultrasound Image Segmentation and Classification of Benign and Malignant Thyroid Nodules on the Basis of Deep Learning

This study aimed to investigate the effect of an image denoising algorithm based on weighted low-rank matrix restoration on thyroid nodule ultrasound images. A total of 1000 original ultrasound image data sets of thyroid nodules were selected as the study samples. The nodule segmentation data set of thyroid ultrasound region of interest (ROI) images was drawn and acquired. By introducing multiscale features and an attention mechanism to optimize the U-Net model, an ultrasound image segmentation model (F-U-Net) was constructed. The performance of the traditional U network model and full convolutional neural network model (FCN) was analyzed and compared by simulation experiments. The results showed that the dice coefficient, accuracy, and recall of the improved loss function in this study were significantly higher than those of the traditional cross entropy loss function and dice coefficient loss function, and the differences were statistically significant (P &lt; 0.05). The Dice coefficient, accuracy, and recall of the F-U-net model were significantly higher than those of the traditional FCN model and U-net model (P &lt; 0.05). The diagnostic sensitivity, specificity, accuracy, and positive predictive value of the F-U-net model for benign and malignant thyroid nodules were significantly higher than those of the FCN model and U-net model (P &lt; 0.05). In summary, the proposed F-U network can effectively process the ultrasound images of thyroid nodules, improve the image quality, and help to improve the diagnostic effect of benign and malignant thyroid nodules. It provides a data reference for segmentation and reconstruction of benign and malignant ultrasound images of thyroid nodules.

Method for Carotid Artery 3-D Ultrasound Image Segmentation Based on CSWin Transformer.

Precise segmentation of carotid artery (CA) structure is an important prerequisite for the medical assessment and detection of carotid plaques. For automatic segmentation of the media-adventitia boundary (MAB) and lumen-intima boundary (LIB) in 3-D ultrasound images of the CA, a U-shaped CSWin transformer (U-CSWT) is proposed. Both the encoder and decoder of the U-CSWT are composed of hierarchical CSWT modules, which can capture rich global context information in the 3-D image. Experiments were performed on a 3-D ultrasound image data set of the CA, and the results indicate that the U-CSWT performs better than other convolutional neural network (CNN)-based and CNN-transformer hybrid methods. The model yields Dice coefficients of 94.6 ± 3.0% and 90.8 ± 5.1% for the MAB and LIB in the common carotid artery (CCA) and 92.9 ± 4.9% and 89.6 ± 6.2% for MAB and LIB in the bifurcation, respectively. Our U-CSWT is expected to become an effective method for automatic segmentation of 3-D ultrasound images of CA.

GUDU: Geometrically-constrained Ultrasound Data augmentation in U-Net for echocardiography semantic segmentation

Echocardiography is a very important medical examination that helps in the computation of critical heart functions. Boundary identification, segmentation and estimation of the volume of key parts of the heart, especially the left ventricle, is a difficult and time-consuming process, even for the most experienced cardiologists. In recent years, research has focused on the automatic segmentation of heart through artificial intelligence techniques and especially with the use of deep learning. Our work is part of this framework.We implemented an ensemble of convolutional neural networks based on the U-net architecture, trained it using a public dataset of cardiac ultrasound images, and combined the outcomes to extract the areas of the left ventricle, myocardium and left atrium. In order to optimize the training process, we have developed a significant data augmentation method based on medical practice. Furthermore, we extended the Dice loss function by imposing additional mandatory anatomical constraints. An ablation study highlights the contribution of each of our proposed modules.The evaluation of our method showed an overall improvement in segmentation accuracy but also in the estimation of clinical metrics. Specifically, using the Dice coefficient for geometric metrics, we achieved for the epicardium a score of 0.96 and 0.955 for the end-diastolic and end-systolic phase respectively. For the clinical metrics of the left ventricle volume, the Pearson correlation coefficient was used where our method gave 0.977, 0.981, 0.897 for the end-diastolic, end-systolic phase and ejection fraction respectively. Scores which up until the writing of this article outperform competitive methods.

DGANet: A Dual Global Attention Neural Network for Breast Lesion Detection in Ultrasound Images.

Deep learning-based breast lesion detection in ultrasound images has demonstrated great potential to provide objective suggestions for radiologists and improve their accuracy in diagnosing breast diseases. However, the lack of an effective feature enhancement approach limits the performance of deep learning models. Therefore, in this study, we propose a novel dual global attention neural network (DGANet) to improve the accuracy of breast lesion detection in ultrasound images. Specifically, we designed a bilateral spatial attention module and a global channel attention module to enhance features in spatial and channel dimensions, respectively. The bilateral spatial attention module enhances features by capturing supporting information in regions neighboring breast lesions and reducing integration of noise signal. The global channel attention module enhances features of important channels by weighted calculation, where the weights are decided by the learned interdependencies among all channels. To verify the performance of the DGANet, we conduct breast lesion detection experiments on our collected data set of 7040 ultrasound images and a public data set of breast ultrasound images. YOLOv3, RetinaNet, Faster R-CNN, YOLOv5, and YOLOX are used as comparison models. The results indicate that DGANet outperforms the comparison methods by 0.2%-5.9% in total mean average precision.

Contrast U-Net driven by sufficient texture extraction for carotid plaque detection.

Ischemic heart disease or stroke caused by the rupture or dislodgement of a carotid plaque poses a huge risk to human health. To obtain accurate information on the carotid plaque characteristics of patients and to assist clinicians in the determination and identification of atherosclerotic areas, which is one significant foundation work. Existing work in this field has not deliberately extracted texture information of carotid from the ultrasound images. However, texture information is a very important part of carotid ultrasound images. To make full use of the texture information in carotid ultrasound images, a novel network based on U-Net called Contrast U-Net is designed in this paper. First, the proposed network mainly relies on a contrast block to extract accurate texture information. Moreover, to make the network better learn the texture information of each channel, the squeeze-and-excitation block is introduced to assist in the jump connection from encoding to decoding. Experimental results from intravascular ultrasound image datasets show that the proposed network can achieve superior performance compared with other popular models in carotid plaque detection.

Transcranial Acoustic Metamaterial Parameters Inverse Designed by Neural Networks.

Objective: The objective of this work is to investigate the mapping relationship between transcranial ultrasound image quality and transcranial acoustic metamaterial parameters using inverse design methods. Impact Statement: Our study provides insights into inverse design methods and opens the route to guide the preparation of transcranial acoustic metamaterials. Introduction: The development of acoustic metamaterials has enabled the exploration of cranial ultrasound, and it has been found that the influence of the skull distortion layer on acoustic waves can be effectively eliminated by adjusting the parameters of the acoustic metamaterial. However, the interaction mechanism between transcranial ultrasound images and transcranial acoustic metamaterial parameters is unknown. Methods: In this study, 1,456 transcranial ultrasound image datasets were used to explore the mapping relationship between the quality of transcranial ultrasound images and the parameters of transcranial acoustic metamaterials. Results: The multioutput parameter prediction model of transcranial metamaterials based on deep back-propagation neural network was built, and metamaterial parameters under transcranial image evaluation indices are predicted using the prediction model. Conclusion: This inverse big data design approach paves the way for guiding the preparation of transcranial metamaterials.

RFI-GAN: A reference-guided fuzzy integral network for ultrasound image augmentation

The Generative Adversarial Network (GAN) is commonly used for medical image augmentation, a method to alleviate the data shortage for downstream tasks. However, traditional GANs is not effective in mining the intrinsic association between samples and data space. Also, the multiple sections and variable scaling ratios of ultrasound images lead to inconsistency among images. The current GAN-based image generation methods cannot obtain the overall anatomical semantic constraints of inconsistent images, thus illogical images are generated. Therefore, a Reference-guided Fuzzy Integral GAN (RFI-GAN) is proposed. First, RFI-GAN implicitly constrains the logicality of the generated images through feature fusion between samples and reference set. These references, a kind of prior knowledge, cover most of the anatomical structures in the dataset. Second, as the general linear fusion takes features as additive and cannot fit the inconsistency of features, we generalize the additivity of features to non-additivity through the Fuzzy Integral, which is used to fusion non-additive measures. Two Fuzzy Integral Modules are proposed to nonlinearly fusion the texture and structure features of ultrasound images under the guidance of references. Finally, we perform experiments on the thyroid ultrasound image dataset, demonstrating the effectiveness of RFI-GAN in terms of generation and downstream classification tasks.

Automatic segmentation of ultrasound images using SegNet and local Nakagami distribution fitting model

Automatic and accurate segmentation of ultrasound (US) images plays a vital role in computer-aided diagnosis of many diseases. However, multiple degradations within US images, such as speckle noise, intensity inhomogeneity, shadows, low contrast, and low signal-to-noise ratio, often cover up the details of imaging tissues and blur the edges of lesions, resulting in limited accuracy for existing segmentation algorithms. To tackle these issues, a coarse-to-fine segmentation method combining deep learning with an active contour model (ACM) is proposed in this paper. At the first stage, a SegNet is trained and employed to predict the segmentation mask because it not only has a superior boundary delineation capability, but also has high efficiency in terms of memory and computational time during inference. Due to the multiple degradations of US images, the unavailability of large-scale US image datasets, and the loss of spatial information during downsampling, the SegNet may only infer the coarse object boundaries, especially for the tissue with blurred edges. At the second stage, the segmentation mask of the SegNet is refined to locate the accurate object boundaries by a novel ACM termed the local Nakagami distribution fitting (LNDF) model, which takes advantage of discrepancies between Nakagami distributions of different tissues around the boundaries within US images. Experimental results on two US image datasets demonstrate that, compared with some state-of-the-art segmentation methods, our proposed method achieves the highest segmentation accuracy at the cost of acceptable running time, and thereby adapts to the automatic segmentation of US images in clinical applications.

Ultrasonographic pathological grading of prostate cancer using automatic region-based Gleason grading network.

The Gleason scoring system is a reliable method for quantifying the aggressiveness of prostate cancer, which provides an important reference value for clinical assessment on therapeutic strategies. However, to the best of our knowledge, no study has been done on the pathological grading of prostate cancer from single ultrasound images. In this work, a novel Automatic Region-based Gleason Grading (ARGG) network for prostate cancer based on deep learning is proposed. ARGG consists of two stages: (1) a region labeling object detection (RLOD) network is designed to label the prostate cancer lesion region; (2) a Gleason grading network (GNet) is proposed for pathological grading of prostate ultrasound images. In RLOD, a new feature fusion structure Skip-connected Feature Pyramid Network (CFPN) is proposed as an auxiliary branch for extracting features and enhancing the fusion of high-level features and low-level features, which helps to detect the small lesion and extract the image detail information. In GNet, we designed a synchronized pulse enhancement module (SPEM) based on pulse-coupled neural networks for enhancing the results of RLOD detection and used as training samples, and then fed the enhanced results and the original ones into the channel attention classification network (CACN), which introduces an attention mechanism to benefit the prediction of cancer grading. Experimental performance on the dataset of prostate ultrasound images collected from hospitals shows that the proposed Gleason grading model outperforms the manual diagnosis by physicians with a precision of 0.830. In addition, we have evaluated the lesions detection performance of RLOD, which achieves a mean Dice metric of 0.815.

Robust co-teaching learning with consistency-based noisy label correction for medical image classification.

Deep neural networks (DNNs) have made great achievements in computer-aided diagnostic systems, but the success highly depends on massive data with high-quality labels. However, for many medical image datasets, a considerable number of noisy labels are introduced by inter- and intra-observer variability, thus hampering DNNs' performance. To address this problem, a robust noisy label correction method with the co-teaching learning paradigm is proposed. The proposed method aims to reduce the effect of noisy labels by correcting or removing them. It consists of two modules. An adaptive noise rate estimation module is employed to calculate the dataset's noise rate, which is helpful to detect noisy labels but is usually unavailable in clinical applications. A consistency-based noisy label correction module aims to detect noisy labels and correct them to reduce the disturbance from noisy labels and exploit useful information in data. Experiments are conducted on the public skin lesion dataset ISIC-2017, ISIC-2019, and our constructed thyroid ultrasound image dataset. The results demonstrate that the proposed method outperforms other noisy label learning methods in medical image classification tasks. It is also evaluated on the natural image dataset CIFAR-10 to show its generalization. This paper proposes a noisy label correction method to handle noisy labels in medical image datasets. Experimental results show that it can self-adapt to different datasets and efficiently correct noisy labels, which is suitable for medical image classification.

An enhanced multiscale generation and depth-perceptual loss-based super-resolution network for prostate ultrasound images

Computer-aided diagnosis of prostate ultrasound images is gradually being combined with deep learning to help detect and treat diseases. However, ultrasound images of the prostate have problems such as low resolution and unbalanced categories. In addition, the current image classification algorithms have difficulty with classification network performance due to insufficient data volume. To solve these problems, inspired by CycleGAN, we propose an enhanced multiscale generation and depth-perceptual loss-based super-resolution (SR) network for prostate ultrasound images (EGDL-CycleGAN). We study and improve the generative network and perceptual loss of CycleGAN. In this work we achieve multiscale feature extraction through an improved generator, and utilize full-scale skip connections between encoder and decoder to capture fine-grained details and coarse-grained semantics at full scale. This effectively improves the performance of the generative network and makes the reconstruction effect better. We also use the residual structure for deep extraction of features to obtain perceptual loss, and add this to the network loss function for training the model. This enables the model to learn the global and local differences between the real and generated images. This approach pays more attention to the edge information and spatial information of the image, and provides relevant spatial information feedback to the generator to improve the generator’s ability to perceive consistent super-resolution. The method can enhance the prostate ultrasound image dataset and provide rich images for the next step in intelligence-assisted classification and diagnosis of prostate cancer ultrasound images. The evaluation of peak signal-to-noise ratio/structural similarity and visual effects against the benchmark of our datasets illustrates that our proposed approach is effective and superior to the bicubic classic image SR reconstruction algorithm, the SRGAN perception-driven method and the CycleGAN method applied to ultrasound images. In addition, the method of using the original dataset combined with the SR reconstruction image dataset can effectively improve the accuracy of the classification network in intelligence-assisted classification diagnosis of prostate cancer ultrasound images. In EfficientNetV2 the accuracy is improved from 0.843 to 0.867 and in Swin Transformer the accuracy is improved from 0.893 to 0.917.

BUViTNet: Breast Ultrasound Detection via Vision Transformers.

Convolutional neural networks (CNNs) have enhanced ultrasound image-based early breast cancer detection. Vision transformers (ViTs) have recently surpassed CNNs as the most effective method for natural image analysis. ViTs have proven their capability of incorporating more global information than CNNs at lower layers, and their skip connections are more powerful than those of CNNs, which endows ViTs with superior performance. However, the effectiveness of ViTs in breast ultrasound imaging has not yet been investigated. Here, we present BUViTNet breast ultrasound detection via ViTs, where ViT-based multistage transfer learning is performed using ImageNet and cancer cell image datasets prior to transfer learning for classifying breast ultrasound images. We utilized two publicly available ultrasound breast image datasets, Mendeley and breast ultrasound images (BUSI), to train and evaluate our algorithm. The proposed method achieved the highest area under the receiver operating characteristics curve (AUC) of 1 ± 0, Matthew's correlation coefficient (MCC) of 1 ± 0, and kappa score of 1 ± 0 on the Mendeley dataset. Furthermore, BUViTNet achieved the highest AUC of 0.968 ± 0.02, MCC of 0.961 ± 0.01, and kappa score of 0.959 ± 0.02 on the BUSI dataset. BUViTNet outperformed ViT trained from scratch, ViT-based conventional transfer learning, and CNN-based transfer learning in classifying breast ultrasound images (p &lt; 0.01 in all cases). Our findings indicate that improved transformers are effective in analyzing breast images and can provide an improved diagnosis if used in clinical settings. Future work will consider the use of a wide range of datasets and parameters for optimized performance.

Deep Learning Technology Applied to Medical Image Tissue Classification.

Medical image classification is a novel technology that presents a new challenge. It is essential that pathological images are automatically and correctly classified to enable doctors to provide precise treatment. Convolutional neural networks have demonstrated their effectiveness in classifying images in deep learning, which may have dozens or hundreds of layers, to illustrate the relationship between them in terms of their different neural network features. Convolutional layers consisting of small kernels take weights as input and guide them through an activation function as output. The main advantage of using convolutional neural networks (CNNs) instead of traditional neural networks is that they reduce the model parameters for greater accuracy. However, many studies have simply been focused on finding the best CNN model and classification results from a single medical image classification. Therefore, we applied a common deep learning network model in an attempt to identify the best model framework by training and validating different model parameters to classify medical images. After conducting experiments on six publicly available databases of pathological images, including colorectal cancer tissue, chest X-rays, common skin lesions, diabetic retinopathy, pediatric chest X-ray, and breast ultrasound image datasets, we were able to confirm that the recognition accuracy of the Inception V3 method was significantly better than that of other existing deep learning models.

Automated BI-RADS classification of lesions using pyramid triple deep feature generator technique on breast ultrasound images.

Ultrasound (US) is an important imaging modality used to assess breast lesions for malignant features. In the past decade, many machine learning models have been developed for automated discrimination of breast cancer versus normal on US images, but few have classified the images based on the Breast Imaging Reporting and Data System (BI-RADS) classes. This work aimed to develop a model for classifying US breast lesions using a BI-RADS classification framework with a new multi-class US image dataset. We proposed a deep model that combined a novel pyramid triple deep feature generator (PTDFG) with transfer learning based on three pre-trained networks for creating deep features. Bilinear interpolation was applied to decompose the input image into four images of successively smaller dimensions, constituting a four-level pyramid for downstream feature generation with the pre-trained networks. Neighborhood component analysis was applied to the generated features to select each network's 1,000 most informative features, which were fed to support vector machine classifier for automated classification using a ten-fold cross-validation strategy. Our proposed model was validated using a new US image dataset containing 1,038 images divided into eight BI-RADS classes and histopathological results. We defined three classification schemes: Case 1 involved the classification of all images into eight categories; Case 2, classification of breast US images into five BI-RADS classes; and Case 3, classification of BI-RADS 4 lesions into benign versus malignant classes. Our PTDFG-based transfer learning model attained accuracy rates of 79.29%, 80.42%, and 88.67% for Case 1, Case 2, and Case 3, respectively.