Brief Review — BUSIS: A Benchmark for Breast Ultrasound Image Segmentation
BUSIS: A Benchmark for Breast Ultrasound Image Segmentation,
BUSIS, by Harbin Institute of Technology, University of Idaho, Utah State University, Medical College of Qingdao University, and Hebei Medical University,
2022 MDPI J. Healthcare, Over 10 Citations (Sik-Ho Tsang @ Medium)
Biomedical Image Segmentation
2015 … 2021 [Expanded U-Net] [3-D RU-Net] [nnU-Net] [TransUNet] [CoTr] [TransBTS] [Swin-Unet] [Swin UNETR] [RCU-Net] [IBA-U-Net] [PRDNet] [Up-Net] [SK-Unet] 2022 [UNETR] [Half-UNet]
==== My Other Paper Readings Also Over Here ====
- Breast UltraSound Image Segmentation (BUSIS) dataset is proposed, which consists of 562 breast ultrasound images.
- 16 state-of-the-art segmentation methods are evaluated.
- Breast UltraSound Image Segmentation (BUSIS) Dataset
- In previous works, most approaches were evaluated by using private datasets and different quantitative metrics, as above.
- This make the objective and effective comparisons among the methods quite challenging.
2. Breast UltraSound Image Segmentation (BUSIS) Dataset
- The BUS image dataset has 562 images among women between the ages of 26 to 78 years. They are acquired using multiple ultrasound devices.
- 4 experienced radiologists are involved in the ground truth generation. There are in total 4 steps:
- Each of the 3 experienced radiologists delineate each tumor boundary manually, and 3 delineation results are produced for each BUS image.
- All pixels inside/on the boundary are viewed as tumor region, outside pixels as background. Majority voting is used to generate the preliminary result for each BUS image.
- The senior expert reads each BUS image and refers to its corresponding preliminary result to decide if it needs any adjustment.
- Finally, the tumor pixel is labelled as 1 and the background pixel is labelled as 0. A binary and uncompressed image is generated as the ground truth for each BUS image.
3.1. Looseness Ratio (LR) for Semi-Automatic Segmentation
- For semi-automatic segmentation approaches, the segmentation performances are evaluated using the same set of ROIs and evaluate the sensitivity of the methods to ROIs with different looseness ratio (LR) defined by:
- where BD0 is the size of the bounding box of the ground truth and is used as the baseline, and BD is the size of an ROI containing BD0.
- 10 groups of ROIs with different LRs are generated automatically using the approach described in . 4 sides of an ROI are moved toward the image borders to increase the looseness ratio. The LR of the first group is 1.1; and the LR of each of the other groups is 0.2 larger.
- Under different LRs, different metrics are evaluated.
3.2. Area Error Metrics
- The area error metrics include the true positive ratio (TPR), false positive ratio (FPR), Jaccard index (JI), dice similarity coefficient (DSC), and area error ratio (AER):
- where Am is the pixel set in the tumor region of the ground truth, Ar is the pixel set in the tumor region generated by a segmentation method, and |.| indicates the number of elements of a set.
3.3. Boundary Error Metrics
- Furthermore, Hausdorf error (HE) and mean absolute error (MAE) are used to measure the worst possible disagreement and the average agreement between two boundaries, respectively.
- Let Cm and Cr be the boundaries of the tumors in the ground truth and the segmentation result, respectively. HE is defined as:
- where x and y are the points on the boundaries Cm and Cr, respectively; and d(z, C) is the minimum distance between a point z and all points on a boundary C as:
- where ||z-k|| is the Euclidean distance between points z and k. MAE is defined by:
3.4. Semi-Automatic Segmentation Approach Benchmarking
- 10 ROIs were generated automatically for each BUS image, with an LRs range from 1.1 to 2.9 (step size is 0.2). In total, 5620 ROIs were generated for the entire BUS dataset.
Five metrics (average JI, DSC, AER, HE, and MAE) reach their optimal values at the LR of 1.9.
3.5. Fully Automatic Segmentation Approach Benchmarking
- The performance of 14 fully automatic approaches is benchmarked.
- The average JIs of all deep learning approaches are above 0.8 except FCN-AlexNet; and  achieved the best average JI performance.
But authors also mentioned that deep learning approaches have limitations in segmenting small breast tumors .