Summer 2024 Capstone Project

Machine Learning Model for Classification of Fertilized and Unfertilized Xenopus Frog Embryos in Image Analysis.

In the Summer of 2024, the DeVision team at Louisiana State University (LSU) collaborated with the Aquatic Germplasm and Genetic Resources Center (AGGRC) and the Marine Biological Laboratory (MBL) to utilize a Stardist-based machine learning model for classifying and counting Xenopus laevis (African Clawed Frog) embryos in various developmental stages. The primary objective was to distinguish fertilized embryos from unfertilized ones using high-resolution images and annotated datasets. Stardist, a neural network designed for segmenting biological image data, was used due to its ability to detect round-shaped objects, such as frog embryos. The model achieved a 98.89\% accuracy in counting embryos and a classification accuracy of 93.09\% for unfertilized embryos and 84.92\% for fertilized embryos. The model's accuracy was highest for high-resolution images with minimal embryo overlap, while performance decreased with low-resolution or overlapping images.

To make the model accessible to biological researchers, a user-friendly graphical user interface (GUI) was developed, allowing users to upload images, classify, and count embryos based on model predictions. Future work will focus on refining the model to classify additional developmental stages and enhancing its robustness with more diverse datasets. The entire project, including the dataset, model, and GUI, is available on GitHub for use by AGGRC and other collaborators. The DeVision team acknowledges guidance from Professors Peter Wolenski and Nadejda Drenska, along with support from the LSU Mathematics Department, AGGRC, and MBL.

Counting Oysters with Stardist: Enhancing Marine Data Collection through Deep Learning.

In the Summer of 2024, the Oyster team at Louisiana State University (LSU) collaborated with Louisiana Sea Grant and Dr. Sarah Bodenstein to develop a machine learning-based tool for automating the counting of oyster seeds in images. This project utilized the Stardist instance segmentation model, which is highly effective in detecting irregular shapes like oyster seeds. The dataset, comprising oyster seeds in three size categories, was annotated using Fiji and LabKit, and augmented with flips, rotations, and intensity rescaling. The model was trained on 500 epochs using a U-net convolutional neural network architecture and achieved 91\% accuracy when predicting oyster seeds in images with darker backgrounds. A graphical user interface (GUI) was developed to allow researchers to upload images and predict oyster seed counts efficiently.

Future work will focus on expanding the dataset to include a greater variety of oyster seed sizes and improving the GUI to support quantification across all size categories. The team also plans to publish research findings to detail the methodology and outcomes. Special thanks go to the Department of Mathematics, Dr. Peter Wolenski, Dr. Nadejda Drenska, Elizabeth M. Robinson from Louisiana Sea Grant, and Dr. Sarah Bodenstein for their guidance and collaboration. All team members contributed to the successful development of this project, which serves as an important step towards enhancing marine data collection using deep learning technologies.

Predicting Appendicular Lean Mass using Supervised and Semi-Supervised Machine Learning Algorithms.

Bridging the Gap: Effective Training on Synthetic Images for Classification of Xenopus laevis Embryo.