Fig 1.
(a) Schematic of bregma and lambda. (b) A subject with a clear view of bregma and lambda. (c) A subject with a obscured view of bregma and lambda.
Fig 2.
(a) A rodent image example. (b) The skull region of the rodent image with bounding box (the blue rectangle) and bregma/lambda masks (the red/green circles) manually labeled. In the red and green circles, the closer a pixel is to bregma/lambda point, the larger the value (i.e., brighter) is assigned to the labeled pixel. Therefore, the edges of the two circles are dark while the center is bright. The bounding box is used to train the faster-rcnn in stage one, and bregma/lambda masks are used to train the FCN in stage two.
Fig 3.
The two-stage localization framework.
In stage one, the localization framework detects the skull area in the low-resolution image using faster-rcnn. In stage two, the localization framework segments bregma and lambda from the full-resolution skull area image using FCN, and the coordinates of bregma and lambda are determined as the maximum value of the remaining segmented area.
Fig 4.
The fully convolutional network.
(a) Detailed structure of the FCN used in stage two of the localization framework. Blue slices represent convolution layers, and orange slices represent bottleneck layers. (b) Detailed structure of one bottleneck layer. Input and output channel numbers are nin and nout, respectively.
Table 1.
The composition of training (6000 images), validation (20 images), and test (13 images) datasets.
Table 2.
Localization errors of different standard deviation labels.
Fig 5.
Mean and standard deviation of training loss and validation loss in 50 epochs across the 4-fold cross validation shown as error bars.
Fig 6.
Examples of input image, ground truth, and the result of the stage two.
Fig 7.
Localization results of the localization framework for 13 testing images.
The left panel shows the localization error of bregma, and the right panel shows the localization error of lambda. Every black dot represents a localization result for a rodent image, and the red cross represents the ground truth position of bregma and lambda.
Table 3.
The comparison of the localization error between our results and humans.
Table 4.
The comparison between the two-stage and the end-to-end approaches.