Fig 1.
(a) and (b) were obtained from Breast Ultrasound Lesions Dataset(Dataset B) [13]. (c) and (d) were acquired from Gelderse Vallei Hospital in Ede, the Netherlands [14]. (e) and (f) were obtained from the Imaging Department of the First Affiliated Hospital of Shantou University. It can be seen from these six figures that there are obvious differences between the tumor morphology and the surrounding tissues.
Fig 2.
Ultrasound breast tumor segmentation based on the RDAU-NET model.
Here (a1)—(c1) were obtained from Dataset B [13]. (a2), (b2), (c2) are gold standard and (a3), (b3), (c3) are the results of the RDAU-NET model segmentation.
Fig 3.
Ultrasound breast tumor segmentation based on the RDAU-NET model.
Here(d1)was obtained from Dataset B [13] and (e1) and (f1) were acquired from the Imaging Department of the First Affiliated Hospital of Shantou University. Also(d2), (e2) and (f2) are gold standard and (d3), (e3), (f3) are the results of the RDAU-NET model segmentation.
Fig 4.
The numbers above the boxes (green) indicate that the size of the input along with the number of channels. For example, 128x128 1 indicate the input resolution and the number of channels respectively. The blue box represents the outputs from Attention Gate module.
Fig 5.
Residual units of encoder and decoder pipeline.
(a) Residual units of encoder pipeline. Here w, h, and b represent the width, height, and channels of the input feature map, respectively. BN is batch normalization. Relu is an activation function and n is the number of filters. In the encoding process, the values of n are 64,128,256,512 and 512 for layers 2,3,4,5,6 respectively. (b)Residual units of decoder pipeline. Here values of n are 512, 256, 128, 64 and 32 for layers 5,4,3,2,1 respectively.
Fig 6.
Illustration of receptive fields for r = 1 and r = 2.
(a) and (b) illustrate the visual field of a 3x3 convolution kernel with r = 1 and r = 2 respectively. When r = 2, though the kernel parameters remain the same, the receptive field has increased to 7x7 (shown as the orange and blue parts in(b)) when compared to traditional convolution (r = 1, as shown in the blue part of (a)). Therefore the dilation process increases the size of the receptive field and compensates for the subsampling.
Fig 7.
Illustration of the dilated convolution module.
Here the dilation parameter r = 2, the stride size S = 1, the input feature map is 4x4, kernel size is 3x3 and receptive filed N is 7x7. After processing using dilated convolution, the size of the original feature map remains the same but the receptive fields increases while keeping the parameters of the model intact.
Fig 8.
Schematic diagram of the Attention Gate (AG).
Fig 9.
Data augmentation illustrating horizontal flipping to expand the training dataset.
Table 1.
Definition of the abbreviations.
Table 2.
The formula of performance measure.
Fig 10.
Sample image from Dataset B [13].
(a) image of malignant invasive ductal carcinoma. (b) Gold standard. (c-f) are the results of segmentation for input sizes are 64x64, 96x96, 128x128, and 256x256 respectively.
Table 3.
Quantitative evaluation of BUS images of different input sizes.
Fig 11.
Segmentation outputs for the BUS images from the test dataset.
The test dataset was obtained from Dataset B. Fig 11(a—d) illustrate the results for test images obtained from Dataset B. (a1), (b1), (c1), (d1) are the gold standard. (a2)—(a12), (b2)—(b12), (c2)—(c12), (d2)—(d12) are the segmentation results from RDAU-NET, FCN8s, FCN16s, SegNet, U-Net, Residual U-Net, Squeeze U-Net, Dilated U-Net, RAU-NET, DAU-NET, RDU-NET respectively.
Fig 12.
Segmentation outputs for the BUS images from the test dataset.
The test dataset was obtained from Dataset B. Fig 12(e—h) illustrate the results for test images obtained from Dataset B. (e1), (f1), (g1), (h1) are the gold standard. (e2)—(e12), (f2)—(f12), (g2)—(g12), (h2)—(h12) are the segmentation results from RDAU-NET, FCN8s, FCN16s, SegNet, U-Net, Residual U-Net, Squeeze U-Net, Dilated U-Net, RAU-NET, DAU-NET, RDU-NET respectively.
Fig 13.
Segmentation outputs for the BUS images from the test dataset.
The test dataset was obtained from the Imaging Department of the First Affiliated Hospital of Shantou University. Fig 13(i—L) represents the outputs for the test images acquired from Imaging Department of the First Affiliated Hospital of Shantou University. (i1), (j1), (k1) and (L1) are the gold standard. (i2)—(i12), (j2)—(j12), (k2)—(k12) and (L2)—(L12) are the segmentation results from RDAU-NET, FCN8s, FCN16s, SegNet, U-Net, Residual U-Net, Squeeze U-Net, Dilated U-Net, RAU-NET, DAU-NET, RDU-NET respectively.
Fig 14.
RDAU-NET performance indicators for training, validation.
Here the plots (a—d) represent the performance metrics during training and validation.
Fig 15.
RDAU-NET performance indicators for training, validation.
Here the plots (e—h) represent the performance metrics during training and validation.
Fig 16.
RDAU-NET performance indicators for training, validation, and testing.
Here the plots (i) represent the performance metrics during training and validation and plots (j and k) specify the performance during testing: Fig 16(j) denotes ROC curve and AUC with respect to True Positive Rate and False Positive Rate and Fig 16(k) illustrate the AUC in relation to Precision and Recall.
Table 4.
Quantitative segmentation results for different models based on the testing dataset.