A VGG attention vision transformer network for benign and malignant classification of breast ultrasound images

Abstract

Purpose Breast cancer is the most commonly occurring cancer worldwide. The ultrasound reflectivity imaging technique can be used to obtain breast ultrasound (BUS) images, which can be used to classify benign and malignant tumors. However, the classification is subjective and dependent on the experience and skill of operators and doctors. The automatic classification method can assist doctors and improve the objectivity, but current convolution neural network (CNN) is not good at learning global features and vision transformer (ViT) is not good at extraction local features. In this study, we proposed a visual geometry group attention ViT (VGGA‐ViT) network to overcome their disadvantages. Methods In the proposed method, we used a CNN module to extract the local features and employed a ViT module to learn the global relationship among different regions and enhance the relevant local features. The CNN module was named the VGGA module. It was composed of a VGG backbone, a feature extraction fully connected layer, and a squeeze‐and‐excitation block. Both the VGG backbone and the ViT module were pretrained on the ImageNet dataset and retrained using BUS samples in this study. Two BUS datasets were employed for validation. Results Cross‐validation was conducted on two BUS datasets. For the Dataset A, the proposed VGGA‐ViT network achieved high accuracy (88.71±$\ \pm \ $1.55%), recall (90.73±$\ \pm \ $1.57%), specificity (85.58±$\ \pm \ $3.35%), precision (90.77±$\ \pm \ $1.98%), F1 score (90.73±$\ \pm \ $1.24%), and Matthews correlation coefficient (MCC) (76.34±7$\ \pm \ 7$3.29%), which were better than those of all compared previous networks in this study. The Dataset B was used as a separate test set, the test results showed that the VGGA‐ViT had highest accuracy (81.72±$\ \pm \ $2.99%), recall (64.45±$\ \pm \ $2.96%), specificity (90.28±$\ \pm \ $3.51%), precision (77.08±$\ \pm \ $7.21%), F1 score (70.11±$\ \pm \ $4.25%), and MCC (57.64±$\ \pm \ $6.88%). Conclusions In this study, we proposed the VGGA‐ViT for the BUS classification, which was good at learning both local and global features. The proposed network achieved higher accuracy than the compared previous methods.

A deep supervised transformer U‐shaped full‐resolution residual network for the segmentation of breast ultrasound image

“…38 This dataset had 980 samples obtained using an iU22 xMA-TRIX scanner (Philips, USA) or a LOGIQ E9 scanner (GE, USA), which included 595 malignant samples and 385 benign samples. 39 The other dataset (dataset B)…”

BUS‐BRA: A breast ultrasound dataset for assessing computer‐aided diagnosis systems

“…Nevertheless, a recent tendency is to classify lesions in BI-RADS assessment categories, which relates the outcome to a malignancy risk and is aligned with the BI-RADS-based reports. In this context, the BUS-BRA dataset can be used to develop and evaluate CAD systems for classifying pathology classes 66,67 and BI-RADS categories. 12,13 It is worth mentioning that the main limitation of BI-RADS annotations is that they are human-dependent and biased toward the specialist's expertise, becoming critical when having a single specialist, as in our project herein.…”

A Hybrid Workflow of Residual Convolutional Transformer Encoder for Breast Cancer Classification Using Digital X-ray Mammograms

“…Recently, a set of researchers used such a technique to recognize benign from malignant cases, i.e., Gheflati et al examined the performance of pure and hybrid pre-trained vision transformer models based on two breast ultrasound datasets [ 28 ], demonstrating the importance of involving the Vision Transformer technique for automatically detecting breast masses in ultrasonography. Another work used a CNN module to extract local features while a ViT module was employed to identify the global features among several areas and improve the relevant local features [ 29 ]. The hybrid model achieved a high precision of 90.77%, recall of 90.73%, specificity of 85.58%, and F1 score of 90.73%.…”