Dataset of breast ultrasound images

1The data collected at baseline include breast ultrasound images among women in ages between 25 and 75 years old. This data was collected in 2018. The number of patients is 600 female patients. The dataset consists of 780 images with an average image size of 500 × 500 pixels. The images are in PNG format. The images are categorized into three classes, which are normal, benign, and malignant. The number of images in each class is shown in Table 1. The data samples are illustrated in Fig. 1. Samples of original images and the images after preprocessing are shown in Fig. 2 and Fig. 3, respectively. Furthermore, each image has its own ground truth (mask image) as shown in Fig. 4.

2Ultrasound (US) images are generally in grayscale. They were collected and stored in a DICOM format at Baheya hospital. The consumed time used to collect and annotate the images is about one year. US dataset is categorized into three classes: normal, benign, and malignant. At the beginning, the number of images collected was 1100. After performing preprocessing to the dataset, the number of images was reduced to 780 images. The original images contain unimportant information not used for mass classification. Moreover, they may affect the output results of the training process. The instruments used in the scanning process are LOGIQ E9 ultrasound system and LOGIQ E9 Agile ultrasound system. These instruments are usually used in top-notch imaging for radiology, cardiac and vascular application. They produce image resolution of 1280*1024. The transducers are 1–5 MHz on ML6-15-D Matrix linear probe. Fig. 2 Illustrates a sample of the original scanned images.

3To make the dataset useful, some tasks should be performed. The data included duplicated images that required to be removed. Furthermore, radiologists from Baheya reviewed and fixed the incorrect annotation. DICOM images were converted to PNG format by using a DICOM converter application. After refining the dataset, the number of US images was reduced to 780 images. The images are categorized into three classes (cases), which are normal, benign, and malignant. All images were cropped to different sizes to remove unused and unimportant boundaries from the images. We used fast photo crop for this task. The image annotation is added to the image name. Special radiologists at Baheya hospital reviewed and checked all images. An example of the refined images is shown in Fig. 3.

4Ground truth (image boundary) is performed to make the ultrasound dataset beneficial. Matlab is used to perform this step. A freehand segmentation is established for each image separately. An example of mask images is shown in Fig. 4. Three folders are created for each type of breast cancer categories. Each folder has the images of its class. The image name includes the name of the class and the number of the image. Furthermore, the name of the masked image has the name as the US images with adding “_mask” to the end name of the image.

5Researchers are mindful of the fact that patients have a right to be protected from public scrutiny of their private lives and illness. To this end, the researcher ensured that the patients and the hospital were adequately informed about the objective of this study. In addition, every patient's data stays unknown and his or her illness states is with the utmost confidentiality.