Introduction

Breast cancer is among the most lethal diseases in females. Every year, breast cancer accounts for 12.5% of all new cancers worldwide. In 2024, approximately 310,720 women and 2,800 men will be diagnosed with breast cancer1. Early detection and diagnosis improve the survival of patients with breast cancer. Thus, the 5-year relative survival rate in the U.S. for all types and stages of breast cancer is 91%, and for localized/early-stage breast cancer, it is 99%2. Moreover, the World Health Organization (WHO) predicts that 19.3 million new cases of cancer will be diagnosed worldwide in 2025.

Generally, clinicians identify suspicious breast lesions after segmentation to diagnose abnormalities. The number of breast cancer patients has increased, so clinicians face difficulties in accurately detecting this disease in a short time. Ultrasound imaging is one of the most common, noninvasive, and inexpensive methods for detecting and characterizing breast disorders. It provides information about the structure of a breast mass. Many features of breast cancer imaging could serve as noninvasive imaging biomarkers to facilitate cancer diagnosis3. The image features extracted from the breast ultrasound images focused on contiguous tissue structures, shapes, borders, and characteristics of surrounding tissues.

Most of the reported results on the classification of breast lesions are based on data collected from within the evaluated tumors. However, in invasive breast cancer, abnormalities may occur in the surrounding tissue. Currently, it is unanimously recognized that the microenvironment surrounding tumors contributes to their formation process4. Breast edema can be caused by either benign or malignant conditions. One predictor of malignancy of breast lesions is so-called peritumoral edema, and a few studies have evaluated the relationship between breast lesion imaging features and cancer occurrence5,6,7,8. On the other hand, the pattern of growth and border irregularities, the penetration degree of the adjacent tissue, and intratumoral tissue-level changes are equally important features for diagnosis.

Despite routine and wide usage, breast ultrasound imaging has limitations, most notably in its low sensitivity and specificity for small masses or solid tumors. Moreover, for an accurate diagnosis, in addition to physician performance, the required number of ultrasound scans to cover the entire breast depends on the breast size9. On average, two or three volumes are acquired for each breast per examination, so large volumes of breast ultrasound images have to be reviewed with a great consumption of time and attention for accurate disease diagnosis. These limitations and the demand for a less time-consuming classification task led to the development of new tools based on artificial intelligence (AI) for breast lesion investigation. Recent advances in AI and medical imaging have led to the widespread use of deep learning (DL) technology, particularly in image processing and classification10,11. This technology enables the automatic extraction of features for breast cancer diagnosis, contributing to a certain extent to improving the accuracy of the diagnosis12. By reducing the number of false positives and negatives, DL networks provide highly accurate breast cancer detection. These methods can be used to diagnose different breast cancer subtypes at various stages by analyzing clinical and test data. To address the limited number of available annotated images, various DL networks pre-trained on large image databases are now available. Transfer learning also overcomes the requirement of machine learning algorithms that training and testing data be in the same feature space. A new framework is now available in which high-performance classifiers are trained using large amounts of data from different domains13,14.

In clinical practice, breast masses are distinguished and classified based on shape, border, and characteristics of the tissue structure. Additionally, calcification shape and lesion edges are important diagnostic clinical features. The significant progress made in radiology and imaging sciences, especially in the use of convolutional neural networks (CNNs) for breast tumor classification tasks, was strongly influenced by the use of transfer learning to reduce the training time and improve the overall performance of DL tasks15. Sirjani et al.16 proposed a novel DL architecture to classify ultrasound-guided breast lesions. They converted InceptionV3 modules to residual inception modules and performed some improvements, such as increasing the number of modules and changing the hyperparameters. A combination of five datasets (three public datasets and two prepared from different imaging centers) was used for training and evaluation. The improved InceptionV3 model classified breast tumors with the following performance metrics in the test group: 0.83, 0.77, 0.8, 0.81, 0.81, 0.18, and 0.77 for precision, recall, F1 score, accuracy, AUC, root mean squared error, and Cronbach’s α, respectively. Montaha et al.17 reported a graph convolutional network (GCN) model for classifying breast tumors as benign or malignant based on graph data. The authors used clinical features/clinical markers such as tumor shape, orientation, margin, and surrounding tissue for classification. The graph attention network (GAT) and GCN were employed for classification. The GCN model outperforms the GAT model in terms of performance. Falconi et al.18 used transfer learning for malignancy classification of breast abnormalities and assessed several pre-trained networks (ResNet50, NASNet, InceptionV3, and MobileNet) when two different preprocessing approaches were employed. When the Otsu thresholding method was not used, ResNet50 returned the best accuracy (0.78). When the Otsu thresholding method was applied, NASNet was the best model, with an accuracy of 0.68. Roslidar et al.19 explored the effectiveness of fine-tuning ResNet101, DenseNet, MobileNetV2, and ShuffleNetV2 models, for breast cancer detection. These models were trained on the ImageNet database and achieved high performance in image classification. The results indicate that ResNet101 and DenseNet achieved 100% accuracy after 10 training epochs, while MobileNetV2 and ShuffleNetV2 required 20 and 30 epochs, respectively. The test results showed that DenseNet correctly classified all the datasets, while ResNet101 and MobileNetV2 performed similarly with 99.6% accuracy, and ShuffleNetV2 showed a slightly lower accuracy at 98%. Heikal et al.20 proposed a customized CNN model that outperforms other pre-trained models for breast cancer detection, including MobileNetV3, EfficientNetB0, VGG16, and ResNet50V2. While the pre-trained models achieved 74–82% accuracy, the customized CNN model achieved an accuracy exceeding 84%. Further optimization performed using gray wolf optimization (GWO) and modified gorilla-troops optimization (MGTO) methods significantly improved the performance of the customized CNN model. With MGTO optimization, the CNN model achieved a remarkable accuracy of 93.13% in just 10 iterations. Kalafi et al.21 employed CNN models to classify benign and malignant breast lesions from two ultrasound imaging datasets. The proposed model with VGG16 features and a fully connected network of only 10 neurons achieved the best performance in the classification task, with 92% and 93% accuracy, respectively. Moon et al.22 developed a computer-assisted diagnostic system (CAD) that combines pre-trained models VGGNet, ResNet, and DenseNet. The developed system was used to classify breast tumors. The model was tested on two breast ultrasound imaging datasets, the BUSI public database and the SNUH private database. For the SNUH dataset, the model obtained an accuracy of 91.00% and an AUC of 0.9697, and for the BUSI dataset, the values were 94.62% accuracy and 0.9711 for AUC scores, respectively. Susilo et al.23 aimed to improve the accuracy of early detection of breast cancer on mammographic images using a CNN. The VGG-16, VGG-19, and ResNet-50 models returned the same accuracy of 86% for the training set. For the validation set, the ResNet-50 model produced the highest level of accuracy (71%) compared to VGG16 (64%) and VGG19 (61%). Dash et al.24 used a CNN to extract patient characteristics at different densities and to distinguish between normal and suspicious areas in mammography images. This study compares the VGG16, MobileNetV2, and ResNet50 models. Lower test accuracies were obtained at age 1 and improved rapidly until age 50, when accuracy values of 97.22%, 98.61%, and 100% were achieved for VGG-16, MobileNetV2, and ResNet50, respectively. Ansar et al.25 reported that the performance of transfer learning models could be impacted when different datasets are used. They investigated transfer learning using a MobileNet architecture for breast cancer classification. Two datasets, the Digital Database for Screening Mammography (DDSM) and the curated breast imaging subset of DDSM (CBIS-DDSM), were used. The authors reported accuracies of 86.8% and 74.5% for DDSM and CBIS-DDSM, respectively. Hanis et al.26 used thirteen pre-trained networks to detect breast anomalies in digital mammograms. They selected an ensemble of ResNet101, ResNet152, and ResNet50V2 models that provided the best performance in classifying suspicious and normal cases. Jabeen et al.27 reported a high accuracy of 98% when implementing a pre-trained deep learning model (DarkNet-53) to extract intricate and high-level features for breast cancer diagnosis. DarkNet-53 is an effective feature extractor that can capture hierarchical representations from medical images, facilitating subtle pattern recognition in patients with breast cancer.

In this study, a method for breast lesion classification using CNNs and transfer learning was proposed. Three pre-trained deep learning models were employed for feature extraction and classification when useful informational content varied through the employed morphological operations. To improve the classification accuracy and model interpretability, the proposed model leverages transfer learning to accelerate the training process. The novelty and uniqueness of the proposed system lie in the approach of considering the features extracted from the tissue inside the tumor (by performing an erosion operation) and from the lesion and surrounding tissue (by performing a dilation operation) for classification. The proposed framework contains two main steps: (i) segmentation of the tumor region followed by erosion and dilation operations, which helps in focusing only on the features extracted from a specific area of the image; and (ii) classification of the extracted lesion region into malignant and benign regions using three pre-trained networks and transfer learning.

The main contributions of this study are summarized as follows:

  • In clinical practice, BI-RADS (Breast Imaging Reporting and Data System) features are difficult for practitioners/radiologists and are time-consuming. To become time-efficient and robust in classification and to capture spatial dependencies, we focused on discrete or localized areas inside and around breast lesions as significant attributes. We investigated the relationship between breast cancer imaging features and inter- and extra-lesional tissue characteristics and the impact of these features on refining the performance of DL classification.

  • Preprocessing the breast images before application to the models. Morphological erosion and dilation generate inter- and extra-lesional tissue areas.

  • Three pre-trained deep neuronal networks (MobileNetV2, VGG16, and EfficientNetB7) and transfer learning give us the advantage of testing the proposed system in a different feature space for a robust classification task. We improve the performance of the pre-trained model by fine-tuning these pre-trained models on an unseen dataset.

  • The proposed system uses the BUS-BRA database (two classes/benign and malignant) for training (70%) and testing (30%) and the unseen BUSI database (two classes/benign and malignant) for testing.

The results indicate that the proposed system, which consists of a preprocessing method, DL pre-trained models, and a transfer learning strategy, outperforms the existing breast cancer detection algorithms reported in the literature.

Results

The proposed framework contains two main steps: (i) segmentation of the tumor region followed by erosion and dilation operations, which helps in focusing only on the features extracted from a specific area of the image; and (ii) classification of the extracted lesion region as malignant or benign. For this purpose, the efficiency of three pre-trained classifiers was checked. They used four classes of input images: raw images provided by the datasets (coded as raw), segmented images using the ground truth (mask) images provided in the datasets (coded as ROI), and eroded and dilated images (coded as Eroded_ and Dilated_). The performance of the proposed models was evaluated on 30% of the remaining images from the BUS-BRA and for the unseen BUSI dataset.

Transfer learning was applied to three pre-trained deep neuronal networks (MobileNetV2, VGG16, and EfficientNetB7), and the accuracy and area under the curve (AUC) were compared. Table 1 provides a summary of the experimental configurations.

Table 1 Experimental configurations of the proposed system.
Full size table

To assess the performance of the proposed system, the accuracy and AUC acquired on the test datasets are presented in Figs. 1, 2, 3 and 4. These metrics need to be maximized for a robust and performant classification. The data provide an overview of each model’s performance and allow a fair comparison of the classification performance. This information facilitates the choice of the best-performing model.

Fig. 1
figure 1

The accuracy values on test datasets, for all CNN architectures.

Full size image

The accuracy values of the pre-trained CNN models on the test datasets are shown in Fig. 1. When both benign and malignant classes were equally important, the overall analysis highlighted the greater ability of the eroded images to distinguish between benign and malignant lesions in terms of accuracy. The accuracy reflects how the model performs across all the classes. The BUSI dataset yielded noticeably higher accuracy values for all the pre-trained networks, i.e., 0.941, 0.931, and 0.959 for MobileNetV2, VGG16, and EfficientNet-B7, respectively. The preprocessing morphological operations affect the classification accuracy, as the eroded images improve the accuracy by an average of 5% compared to the dilated images. The raw images yielded lower accuracies, whereas the segmented images (_ROI) had comparable performance to the preprocessed images. Thus, for segmented images (ROIs), EfficientNetB7 and MobileNetV2 achieve the best performance results in terms of accuracy, whereas VGG16 has a slightly lower accuracy.

The performance of the DL models for breast lesion classification was also quantified using the area under the receiver operating characteristic (ROC) curve (AUC). A higher AUC indicates better classification performance. Figures 2, 3 and 4 summarize the AUC scores for each model.

Fig. 2
figure 2

Model prediction ROC curves of all CNN architectures when the BUS-BRA dataset was used for training (70%) and testing (30%). (a) raw images; (b) ROI images; (c) eroded images; (d) dilated images.

Full size image
Fig. 3
figure 3

Model prediction ROC curves of all CNN architectures when 70% of the BUS-BRA was used for training and 30% of BUS-BRA and the whole BUSI datasets were used for testing. (a) raw images; (b) ROI images; (c) eroded images; (d) dilated images.

Full size image
Fig. 4
figure 4

The area under the curve (AUC) score for all CNN architectures.

Full size image

In the dilation operation case, two pre-trained neural networks, namely, VGG16 and EfficientNet-B7, demonstrated high diagnostic accuracy in detecting breast cancer, for the BUSI dataset. They achieved AUC values of 1.00 on the unseen BUSI test set and 0.954 and 0.994, respectively, on the BUS-BRA test set. The AUCs of the MobileNetV2 model were 0.8974 (BUS-BRA) and 0.765 (BUSI), which are lower than the performance scores returned by the dilation operation. In the eroded image scenario, MobileNetV2 obtained an AUC of 1.00 for both the BUS-BRA and BUSI test sets. VGG16 and EfficientNet-B7 achieved moderate AUC values. The area under the curve (AUC) decreased by 20.4% (BUS-BRA) and 9.26% (BUSI) for VGG16 and by 19.6% (BUS-BRA) and 9.5% (BUSI) for EfficientNetB7 for eroded images. The raw images (not preprocessed) exhibited the worst performance. This indicates the advantage of morphological operations on the final classification. When morphological operations are not used (i.e., _ROI image dataset), CNNs still perform well because these images retain all the information.

Discussion

The novelty of the proposed system lies in considering the features extracted from the tissue inside the tumor (by performing an erosion operation) and from the lesion and its surrounding tissue (by performing a dilation operation) for classification.

An extensive analysis of the previous studies regarding BC classification indicates that CNNs mostly used the whole image as input. In this case, the computing performance increases considerably. The proposed method provides as inputs resized images to the ROI dimensions and convolution operations integrated into CNN become more computationally efficient. The percentage of ROIs in GT binary images, in eroded and dilated images is,

$$\:p=\frac{1}{N}{\sum\:}_{i=1}^{N}\frac{sum\left({J}_{k}\right)}{n\cdot\:m}$$

where Jk is replaced by ROI areas, eroded, and dilated images; the function sum() summarizes the white pixel that forms the lesion. N is the number of images from the BUSI and BUS-BRA datasets, and n and m are the size of images. The percentage depends on the size of the breast lesion. Table 2 shows the percentage of processed information amount after the morphological operations relative to the processing of the entire image.

Table 2 The percentage of processed information amount.
Full size table

Furthermore, the proposed approach is compared with other state-of-the-art methods in Table 3. The comparison considers the CNN systems as black boxes, without details on their architectures. The present study has demonstrated better results.

Table 3 Comparison between the proposed model and related studies.
Full size table

The erosion and dilation morphological approach has been proposed because, for malignant lesions, there is no clear transition between the tumor and peritumoral area as the border is less defined. Additionally, malignant and benign breast lesions can have varying effects on nearby tissues. Thus, the pattern of growth and border irregularities, the penetration degree of the adjacent tissue, and intratumoral tissue-level changes are equally important features for diagnosis.

To verify the scalability of proposed DL models, we train the models on the BUS-BRA dataset and directly test the models on 30% of the BUS-BRA dataset + BUSI dataset, without any adaptation. The effectiveness of transfer learning depends on the relevance of the pre-trained model to the new task. The unseen BUSI dataset contains the same image categories but is not too similar to the BUS-BRA dataset. The inference has been performed on unseen data belonging to categories presented in the training subset. The worsening of the predictions in the BUS-BRA dataset could be the result of a different image quality. Overall, our accuracy results show that the considered DL architectures perform in much the same way on the unseen dataset. High AUC values are essential for creating an effective model to mitigate the risk of breast cancer and they are provided by both eroded and dilated operations.

This study demonstrated that using pre-trained CNNs and transfer learning in breast ultrasound images can improve classification performance. Additionally, the shortage of labeled and ground truth images can be alleviated by integrating transfer learning into the system. The above analysis indicates that erosion and dilation operations improve classification performance and could be coupled with segmentation in the preprocessing stage to improve the overall performance of the networks. We believe that the narrowing and expansion of studied breast lesions could have a significant impact on diagnosis.

This study has several limitations. Real-world data may be influenced by variations in patient characteristics, imaging equipment, and protocols. The ground truth image information of the target domain is often limited or even unknown. Ground truth images are essential and can be provided by experienced physicians. This lack of information makes it difficult to estimate the differences between classes.

Conclusions

The proposed method establishes a novel approach to breast cancer detection. In the present study, the effectiveness of breast cancer classification was compared based on two types of ultrasound data collected from within the tumor (using the erosion operation) and from the tumor and surrounding tissue (using the dilation operation). In the BUSI dataset and erosion cases, the accuracy was greater for all CNNs than for the dilation case. The AUC values calculated for VGG16 and EfficientNet-B7 reached the highest value of 1.00 for dilation cases. On the other hand, the AUC decreased by 20.4% (BUS-BRA) and 9.26% (BUSI) for VGG16 and by 19.6% (BUS-BRA) and 9.5% (BUSI) for EfficientNetB7 for eroded images. The proposed models demonstrate similar performance on the unseen dataset according to the accuracy results. Besides, the features extracted from the inter- and extra-lesion tissue areas provided high AUC values that are essential for creating an effective model to mitigate the risk of breast cancer. These results suggested differences in the ultrasound features between the tissue inside and surrounding the benign and malignant tumors. However, our study indicated that reducing the analysis of the breast lesion area by erosion had a comparable effect to increasing the analysis area by dilation. In conclusion, erosion and dilation operations improve the classification and overall performance of the networks.

Further studies are necessary to provide deeper insights and investigate these findings. In future work, the proposed system could also be effective at assisting in the diagnosis of other cancerous lesions, such as melanoma or thyroid tumors. Additionally, various AI-powered synthetic data generators could be used to overcome the shortage of datasets that contain ground truth images.

This study is among the few studies demonstrating the greater potential of AI as a second reader/opinion in breast cancer diagnosis.

Materials and methods

To date, most studies have focused on classifying breast cancer using deep learning models and raw images or image-based features, neglecting clinically meaningful features and the optimization process. Generally, most researchers do not pay attention to the preprocessing step and instead focus on the segmentation step, followed by feature extraction.

Detailed view of the proposed system

Figure 5 illustrates the architecture of the proposed framework for breast cancer classification using ultrasound images.

Fig. 5
figure 5

The classification flowchart of breast mass by pre-trained CNNs.

Full size image

This study introduces an approach based on analyzing the tissue inside the tumor and the tissue surrounding it to increase diagnostic efficiency from a classification point of view. Erosion and dilation morphological operations are performed on the original ultrasound images from both databases to facilitate novel feature extraction. The preprocessed images are obtained after morphological operations (using a disk-shaped structuring element with a 10-pixel radius). The eroded images consider only the inter-lesion tissue area diminished with a width of 10 pixels or 2.64 mm. The dilation operation ensures the extraction of features from the lesion tissue and an additional 10-pixel or 2.64 mm wide area of extra-lesional tissue. When morphological operations use a disk-shaped structuring element, two complementary shape effects occur: dilation rounds the convex boundaries and preserves the concave boundaries; conversely, erosion rounds the concave boundaries and preserves the shape of the convex boundaries.

The raw, segmented, and preprocessed images were passed to three pre-trained deep networks, MobileNetV2, VGG16, and EfficientNetB7, for training purposes. Training is performed using the extracted features, transfer learning, and fine-tuning for optimization. The fine-tuning mechanism seeks the best configurations for each pre-trained transfer learning CNN model to perform image classification for datasets never trained on. The training step is performed on 70% of the images in the BUS-BRA dataset. The testing step is performed on the 30% remaining images of the BUS-BRA dataset. Additionally, to evaluate the robustness and generalizability of the proposed system, the testing dataset was constructed using the BUSI database as an unseen image dataset.

Dataset

BUS-BRA and BUSI datasets are used in this study. They are publicly available for academic and research purposes through open-access repositories. The Breast Ultrasound Dataset for Assessing CAD Systems (BUS-BRA)33 dataset contains 1875 images of 1064 female patients collected using four ultrasound scanners34. The images were obtained during studies at the National Institute of Cancer (Rio de Janeiro, Brazil). The dataset includes ground truth images that separate breast images into tumoral and normal lesions. 722 benign and 342 malignant cases were confirmed by biopsy, and the files were saved in .png file format (Fig. 6). The dataset is split into 70% for the training set and 30% for the testing set. The images were randomly assigned to training and test sets.

Fig. 6
figure 6

Sample ultrasound images of the BUS-BRA dataset. The first row depicts benign images. The second row depicts malignant images. (a) Original raw image on the left breast; (b) Ground truth image on the left breast; (c) Original raw image on the right breast; (d) Ground truth image on the left breast.

Full size image

A total of 780 breast ultrasound images in .png format are accessible from the public BUSI image dataset35. The BUSI dataset includes 437 benign, 210 malignant, and 133 normal images36. We have selected only 647 images containing benign and malignant classes while the healthy/normal class is not involved in this study. The images are 8-bit gray-level images with an average size of 500 × 500 pixels. Using a segmentation technique and the MATLAB programming environment, radiologists and computer specialists created a dataset of ground truth images, which are also accessible (Fig. 7). The BUSI dataset is used as an unseen image set for testing.

The images are resized so that the final image measures 160 × 160 px.

Fig. 7
figure 7

Sample ultrasound images of the BUSI dataset. (a) Original raw image for benign lesion; (b) Ground truth image for benign lesion; (c) Original raw image for malignant lesion; (d) Ground truth image for malignant lesion.

Full size image

The accuracy, ROC curves, and AUC performance metrics are utilized to evaluate DL algorithms. Accuracy indicates the proportion of correct predictions. ROC curves correlate the true positive rate and the false positive rate at every possible threshold. AUC ranks the positive higher than the negative across a range of threshold values.

Pre-trained CNN models

The benign and malignant lesion classification pipeline starts with an image preprocessing step, followed by a feature extraction stage using pre-trained models and transfer learning. It ends with the classified process itself. CNNs with transfer learning techniques have become more prevalent in classification tasks. Numerous approaches employ fine-tuning to enhance performance. Transfer learning improves model performance by transferring knowledge from a source domain (i.e., ImageNet) to a related target domain. The goal is to enhance learning in the target domain. During transfer learning, the size of the dataset required for training in the target domain can be small, or in some cases, no information in the training phase is needed. The Keras Applications offer 38 pre-trained CNNs for use. From these CNNs, only the MobileNetV2, VGG16, and EfficientNetB7 models were selected because they were trained on ImageNet and have proven their efficacy in classification tasks. The MobileNetV2 and VGG16 models achieved a classification accuracy of 90.1% and an accuracy of 94.2% on the ImageNet dataset. The ResNet family was not selected because it was surpassed in terms of accuracy, training speed, and parameter efficiency by the selected models. Similarly, the Vision Transformers (ViT) are “data hungry” and do not generalize well when trained on small datasets. ImageNet is quite small for ViTs. Finally, the selected models are fine-tuned. The implementation of the steps involved in individual training, transfer learning, and testing of the selected models is depicted in Fig. 8.

Fig. 8
figure 8

Methodology pipeline depicting the steps involved in individual training, transfer learning, and testing the VGG16, MobileNetV2, and EfficientNetB7 models.

Full size image

Initially, the model was trained using more than one million well-annotated images from the ImageNet dataset. The pre-trained models developed for ImageNet can classify images into 1000 object categories because they learn meaningful feature representations for a wide range of images. The networks consider an image as input and output a label for each object into the image together with the probability of belonging to a category of objects. The convolutional layers extract image features that are further used by the last learnable layer and the final classification layer to classify the input image. These layers combine the extracted features into class probabilities, a loss value, and predicted labels. For selected CNNs, the training process was configured as follows: a base learning rate of 0.0001 and the Adam optimizer based on the Adam algorithm to minimize the loss function. We use the binary cross entropy as the loss function. The models were trained for 15 epochs and a batch size of 32. Data augmentation based on random rotation and horizontal flipping was used to enhance the models’ generalizability.

Transfer learning uses a pre-trained network and learns it for new tasks for better and faster convergence. Some older layers are replaced with new layers that are well adapted to a new dataset to retrain a pre-trained network to classify new images. Moreover, fine-tuning a network via transfer learning is a faster and easier approach for performing new tasks properly using only a small number of training images. The adaptation layer of the deep transfer learning model is formed by a fully connected layer (where the extraction of unique features for classification occurs) that uses the output vector of the fixed feature extractor as input. We noted that the fully connected layers are the most computationally intensive part of a network, so adding one or two layers leads to an exponential increase in the number of parameters. This approach required additional training time to train and provided increased accuracy but could also cause overfitting.

We briefly outline the pre-trained models employed in this study.

VGG16

The VGG16 model is a CNN used for classification37. The VGG16 architecture consists of 16 layers, 13 of which are convolutionary layers and 3 of which are fully connected layers. Additionally, VGG16 incorporates max-pooling layers with a 2 × 2 filter size to downsize the spatial dimensions of the output, allowing effective risk mitigation of overfitting38. The last fully connected layer is adapted to produce classifications between the specific classes. VGG16 can be trained using a learning transfer technique in which the convolution layers of the pre-trained model are used to extract relevant features from breast ultrasound images, followed by fine-tuning to adjust classification patterns to the specific dataset.

MobileNetV2

The MobileNetV2 network was designed to balance accuracy and computational efficiency31,39,40. It consists of inverted residual blocks and helps increase efficiency. The architecture of MobileNetV2 contains a series of convolutional layers, depthwise separable convolutions, inverted residuals, bottleneck designs, linear bottlenecks, and squeeze-and-excitation blocks, which help to maintain important information and avoid losses during network propagation. For image classification purposes, this model uses a large dataset for training. The network can classify new images with high accuracy using the knowledge and features learned during training. Its architecture is suitable for applications such as breast and skin cancer classification.

EfficientNet-B7

EfficientNet-B7 is a CNN consisting of a series of building blocks that follow an innovative scaling architecture composed of a variable number of expansion, convolution, and compression stages41. The network is scaled into three dimensions, depth, breadth, and resolution, which allows the model to learn complex data features with low computational costs. EfficientNet-B7 features a deep and intricate architecture with numerous layers, enabling it to capture detailed information from breast ultrasound images. In terms of ImageNet accuracy, EfficientNetB7 outperforms prior state-of-the-art CNNs42.

Transfer learning and fine-tuning

Training and testing data may not be in the same feature space in real-world applications. Transfer learning is adequate for using knowledge or pre-trained models on a new or similar issue. Instead of training a model from scratch, transfer learning uses a pre-trained model on a vast dataset, such as the standard ImageNet dataset, which has already learned features. The pre-trained model is adapted to solve a specific problem using a smaller and more specific dataset14. This strategy leverages the acquired knowledge during the initial training process and trains it on a more specific task. During fine-tuning, the CNN model starts with pre-trained weights from the initial training, which are subsequently adapted to better suit the new task. Some layers of the grid can be frozen, retaining the general features learned during pre-training, while others can be adjusted to better match the data and requirements of the specific problem43,44,45. The fine-tuning hyperparameters are, the learning rate of 0.001; epochs of 20; layer onwards of 100; and momentum of 0.9.