Semen samples from 62 dogs have been used to build an image database containing 11,464 images of a single sperm obtained using phase-contrast under 1000X magnification. Three specialists are currently classifying the images and building an annotated database to be used to train and test an optimized algorithm for sperm morphology classification. Considerable progress was made on the development of a neural network for sperm evaluation. Two methods to standardize image size proved satisfactory, with image stretching resulting in better accuracy when compared to use of black borders.
Additional possibilities for improving the performance of the implemented model will be investigated, including adjusting training hyperparameters, performing preprocessing alternatives such as contour detection before feeding the neural network, and optimizing the neural network architecture itself. Other techniques, such as regularization, insertion of normalization layers, and dropout, are also suggested to achieve better results.
The Principal Investigator requested and was awarded a no-cost extension to 12/31/2024.