Purchase this article with an account.
Rui Ma, Lili Hao, Yudong Tao, Ximena Mendoza, Mohamed Khodeiry, YUAN LIU, Mei-Ling Shyu, Richard K Lee; Synthetic retinal ganglion cell image generation for deep-learning-based neuronal tracing. Invest. Ophthalmol. Vis. Sci. 2021;62(8):2117.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
While most deep learning-based retinal ganglion cell (RGC) neuronal axon and dendritic tracing algorithms have superior performance over conventional ones, they usually require a large amount of manually annotated training data and do not generalize well when the types of neurons in training and testing datasets are significantly different. We developed and tested an automated deep learning-based neuronal tracing approach that can be applied to trace the axons and dendrites in different types of retinal ganglion cell images that does not require manually traced neuronal images for training.
A data mining algorithm is proposed to learn the noise patterns from raw retinal ganglion cell images. These noise patterns were incorporated into 2,400 digitally reconstructed neuron images downloaded from open source databases of neuronal morphologies to obtain a training dataset of synthetic neuron images. A deep learning-based image segmentation model, ResUNet, was then employed to trace the axons and dendrites in neuronal images. This deep learning model was trained using our synthetic dataset and evaluated on two different testing datasets, including a different synthetic neuron dataset with 300 images and an RGC neuron dataset with 52 images. For the synthetic testing dataset, the digitally reconstructed neurons were utilized as the ground truth images for tracing, while manual tracing results from human experts were considered as the gold standard for the RGC datasets.
Our neuron tracing model can effectively trace the axons and dendrites in both synthetic and RGC neuron images. Quantitatively, our model achieves average foreground, background and overall accuracy of 0.880, 0.994 and 0.993 on the synthetic dataset, 0.757, 0.990 and 0.990 on the RGC dataset, respectively.
The tracing results demonstrate that our neuronal image generation algorithm can accurately produce training data. More importantly, unlike all the other deep-learning-based neuron tracing algorithms, our approach does not require manually traced neuronal image during training and significantly improves raw data analysis times.
This is a 2021 ARVO Annual Meeting abstract.
Input images (left), tracing results from the model (middle), and the ground truth tracing results (right) from the testing data of our synthetic dataset (up) and RGC dataset (down).
This PDF is available to Subscribers Only