The breast lesion detection in ultrasound videos dataset uses a clip-level and video-level feature aggregated network (CVA-Net) and consists of 188 ultrasound videos, of which 113 are labeled malignant and 75 benign. Overall these consist of 25,272 ultrasound images in total with the number of images for each video varying from 28 to 413. 150 videos were used for training, 38 for testing. The primary intended use case would be for computer-aided breast cancer diagnosis, supporting systems to assist radiologists.
Here are more details summarising the approach:
- A novel network: a new state-of-the-art clip-level and video level feature aggregated network (CVA-Net) created to aggregate clip-level temporal features and video-level lesion classification features to fuse and into a prediction classifier. It outperformed existing methods mainly focused on 2D images and or fusing with unlabeled videos.
- The need for increased accuracy contributed to the motivation, given detection challenges due to blurry breast lesion boundaries, inhomogeneous distributions, changeable breast lesion sizes, and positions in dynamic video.
- Each video has a complete scan of the tumor, from where it becomes visible to where it is no longer visible as well as the largest section - all acquired via LOGIQ-E9 and PHILIPS TIS L9-3 ultrasound machines .
- 2 pathologists with 8 years of experience invited to manually annotate the breast lesion rectangles inside each frame and give the corresponding classification.
Second source in addition to separate homepage URL below: https://github.com/jhl-Det/CVA-Net/tree/main/datasets