A practical evaluation of machine learning for classification of ultrasound images of ovarian development in channel catfish (Ictalurus punctatus)

Machine learning is a powerful tool to improve efficiency of industrial processes, but it has not yet been well utilized in aquacultural and hatchery applications. The goal of the present study was to evaluate the feasibility of using a broad array of machine learning approaches (testing of \u3e200 vectorization and model combinations, reporting on 20) to classify ultrasound images spanning annual ovarian development (i.e., from undeveloped to mature) of channel catfish (Ictalurus punctatus). The specific objectives were to: 1) establish dataset preprocessing to standardize image features; 2) develop and train image classification models with deep learning methods; 3) develop and train models with traditional machine learning methods; 4) compare performance of deep learning and traditional methods on two classification problems (2-class and 5-class), and 5) propose insights to deploy models in practical aquaculture applications for research and hatchery use. A total of 931 ultrasound images of catfish ovaries were used to train and evaluate models for a 2-class problem (as a ‘yes’ or ‘no’ answer) to support hormone-injection decisions for spawning management in hatcheries, and a 5-class problem for classifying gonadal development stages for research. By using feature extraction, cropping, dimension reduction, and histogram normalization, a preprocessing method was created to standardize images to develop traditional (i.e., vector input), and deep learning convolutional neural network (CNN) (i.e., image input) approaches. Traditional machine learning models with image classification achieved 100% median accuracy on the 2-class problem (with the models RN-50 and RN-152), and 96% median accuracy for the 5-class problem (with VGG-19 image vectorization). The deep learning approach for the 2-class problem had a median accuracy of \u3e98% for 15models. The 5-class deep learning models produced a steady increase in median accuracy with training net size, achievable through expansion of the dataset. These models can be developed further, but traditional models (using CNN architectures to simply calculate image vectors) outperformed the deep learning approach. These models can be directly applicable to aquaculture in hatcheries and reproductive biology research, in addition to a wide variety of other image-based applications