Fig 1.
The network architecture of RetinaNet.
RetinaNet uses the Feature Pyramid Network (FPN) [34] on top of the convolutional neural network ResNet [35] as a backbone network (a) to generate a rich convolutional feature pyramid (b). The class subnet (c) is for classifying anchor boxes, and the box subnet (d) is for regressing from anchor boxes to ground-truth object boxes. (Lin, Tsung-Yi, et al., 2017 [32].)
Fig 2.
The full mammogram and 25 overlapped sections are used together as an input to our mass detection model.
The inference results of each input image are put together.
Table 1.
Experimental setups.
Fig 3.
True positive rate as a function of the minimum IoU value.
The left figure shows the average precision curves of S-1, S-3, S-5, and S-6 which are results of setups tested on INbreast and the right figure shows the average precision curves of S-2, S-4, and S-6 which are the results of setups tested on GURO.
Fig 4.
Free response operating characteristic (FROC) curves of the results on various operating points, representing the true positive rate (TPR) and the false positive per image (FPPI).
FPPI values on the X-axis are converted to a logarithmic scale. The left figure shows the FROC curves of S-1, S-3, S-5, and S-6, which denote the results on INbreast, and the right figure shows the FROC curves of S-2, S-4, and S-6, which denote the results on GURO.
Fig 5.
X of the curve is the average number of false positives per image, and Y of the curve is the confidence score of our mass detection model.
To show how well our mass detection model works on mammograms without masses, we trained our model on the mammograms containing masses and tested it on the mammograms without masses. The red line denotes the result on INbreast and the blue line denotes the result on GURO. To make it easier to distinguish between similar values, we converted the scale to a log scale, and the confidence score starts at 0.1.
Table 2.
Performance comparison of the mass detection models.
Fig 6.
FROC curves of our mass detection model used in S-1, S-3, and S-6 and the announced results of state-of-the-art mass detection models.
FPPI values on the X-axis are converted to a logarithmic scale.
Fig 7.
Several good and poor cases from the S-1 and S-2 setups.
Mammograms from GURO are shown at the top and mammograms from INbreast are shown at the bottom.