Fig 1.
Zebrafish development and the tail bud classification problem.
A. Schematic drawings of zebrafish embryos at 18, 21, 25 and 30 somite stages. The tail bud region of the 18 somite stage embryo schematic is shown inside the dotted square and below, rotated 90° to the right and zoomed in. Black line on the boxes to help visualise alignment. Posterior to the right, anterior to the left, dorsal up and ventral down. The spinal cord (SC) is shown in cyan, alternate somites (Ss) are shown in different shades of orange, the notochord (NC) is shown in green and the pre-somitic mesoderm (PSM) is shown in pink. B. Maximum projection images of the tail buds of embryos at the 18, 21, 25 and 30 somite stages respectively, stained with Dapi and imaged on a confocal microscope. C. Maximum projection images of the same tail buds as in B, stained for the mRNA of tbxta (green in the posterios PSM and notochord), tbx16 (red in the PSM) and sox2 (blue in the spinal cord) using HCR V.3 and imaged on a confocal microscope.
Fig 2.
A. Same maximum projection image as in Fig 1C (third from the left) of a 25 somite stage embryo’s tail bud stained for the mRNA of tbxta (green in the posterior PSM and notochord), tbx16 (red in the PSM) and sox2 (blue in the spinal cord). B. Same image as in A. but only showing the sox2 (blue) channel. C. Same image as in A. but only showing the tbx16 (red) channel. D. Same image as in A. but only showing the tbxta (green) channel.
Fig 3.
Convolutional neural network architecture.
Convolution and pooling region of the 2D CNN architecture.
Table 1.
Comparison matrix of classification outcomes of 2D and 3D morphological images.
Training accuracy is derived from the highest average accuracy from all epochs and test accuracy, from testing on a subset of data that has not been used during training.
Table 2.
Our data were divided into two data sets, the 2D and the 3D datasets. Within those data sets there were image stained for DAPI, Tbxta, Tbx16 and Sox2. The table shows the total number of images of each kind (Total column), and how many images were used for training (Training column) and Testing (Testing column). In brackets, the number of images per class.
Table 3.
Classification outcomes of the CNNs trained with 3D gene expression image data.
Training accuracy is derived from the highest average accuracy from all epochs and test accuracy, from testing on a subset of data that has not been seen during training.
Table 4.
Classification outcomes of the CNNs trained with 3D gene expression image data.
Training accuracy is derived from the highest average accuracy from all epochs and test accuracy, from testing on a subset of data that has not been seen during training.
Table 5.
Parameter optimisation results for the 3D CNN.
The activation functions and hidden units refer to the parameters used in the first and second layers of the MLP part of the network, respectively. Test accuracy was obtained on a test data set of 24 images unseen by the network during training.
Table 6.
Parameter optimisation results for the 2D CNN.
Hidden units refer to the parameters used in the first and second layers of the MLP part of the network respectively. Test accuracy was obtained on a test data set of 24 images unseen by the network during training.