A convolutional neural network to count the number of neurons in microscope images of cell tissue. Developed in a team as a course final project.
Our neuron-counting model is based off of the Object Classification example notebook from TensorFlow. We adapted the basic functions (build_dataset, build, compile, and train model, and print accuracy metric charts) to our goal of predicting the number of neurons in an image of cell tissue. A focus of this project is using transfer learning to iteratively adapt a pre-existing model to a specialized problem.
Check out this collaborative blog post about our development process and an analysis of our results!
Training data: https://www.kaggle.com/datasets/jamiebergen/neurondata-th-anna1
The attached Jupyter Notebook contains the latest iteration of the neuron-counting model and testing benchmarks. The model can be retrained with new parameters or an expanded dataset. When adding/removing images from the dataset, ensure that data_dir is set to the directory of the original images, and that neuron_labels is an array of the integer counts associated with each image in alphanumeric order. Changing NUM_AUGMENTATIONS will change the number of rotated copies of each image used to supplement the dataset, and will automatically update the image labels to match the supplemented dataset.