Enteritis scoring on distal intestine slides are critical measures for industry study reporting on the effects of alternative feed formulation and dietary ingredients. Limited progress has been made in formalizing these measurements through grading rubrics and rules of thumb (Uran 2008). Current scoring relies on manual grading by a professional histopathologist against a prescribed ordinal scale on several anatomical metrics. Due to the slow, expensive, and inconsistent nature of this human-based approach, there is still significant room for improvement in the cost, accessibility, and robustness of enteritis grading. This proposal aims to train a machine-learning model to rapidly, accessibly, cheaply, and robustly quantify distal enteritis in histology slide images in a bias-free and reproducible manner (Guan, 2022). This technology may be applied to every future project involving distal enteritis evaluation. This project addresses priority 2.1 of the SAA RFP: Understanding gastrointestinal barriers to soy inclusion: specifically, enteritis. We propose that a successfully trained model may achieve enteritis grading at or above the level of a board-certified histopathologist, on the order of milliseconds, at zero cost, in a freely available format which is usable by anyone with access to a gpu-enabled laptop.

(1) Crowdsource 800-1000 commercial salmonid (Atlantic salmon and rainbow trout; approx. equal proportions) distal intestinal slides from multiple industry and research partner laboratories, ranging along the continuum of clinical enteritis from negative (zero inflammation) controls to severe distal enteritis samples, including representatives from all intermediate stages. Efforts will be made to include divergent sources, including domestic and international collaborators (i.e., USA, Norway, Chile, etc.), to improve the variability within the initial training data. Training data variability is expected to increase the robustness and generalizability of the model.
(2) Digitize the obtained slides using a professional-grade slide scanner and assign binary classifications and Uran scores to each digitized slide utilizing two to three independent histopathologists.
(3) Train the ResNet AI model as a binary classifier as proof of concept on these digitized images, to judge each slide as a binary enteritis-positive or enteritis-negative.
(4) Train the model to assign each slide an ordinal grade on each standard anatomical metric, according to the Uran Scale (He, 2016; McCombe, 2021).

(1) A freely accessible, easily transferable lightweight machine-learning program known as a ‘ResNet’ that can grade distal intestine slide images on the Uran Scale and return data on the order of milliseconds per slide at zero cost.
(2) A tool to drastically reduce the cost and turn-around time of enteritis data for every scientific investigation into soy-based diets.
(3) Demonstrate the economic value of machine-learning and artificial intelligence-assisted aquaculture tools to farmers, stakeholders, and economic partners (Sveen 2021; Vo 2021).

Updated June 12, 2024:
Executive Summary
The project, aimed at developing a machine-learning model to score distal intestinal histology slides for soybean meal-induced enteritis, was initial intended to officially commence on April 1, 2024; however, due to administrative delays, the contract was not signed until May 20, 2024 and a formal start date of May 1, 2024 was request. Despite these delays, significant progress was made in these initial months of the project, with a focus on algorithm development and initial model construction.

Q1 Project Activities and Achievements

1. Meetings and Planning:
Dr. Jacob Bledsoe (PI) and Dr. Nathan Redman (coPI) held multiple sessions to outline the construction of the ResNet algorithm and the sourcing of necessary histology slides. A pivotal 2-hour meeting involving the Project PI, Dr. Nathan Redman, and machine learning expert Anita Juhong occurred on May 25th. This meeting focused on determining the optimal parameters for constructing the ResNes analysis network, crucial for the project's success.

2. Development and Scripting of Neural Network:
The team, led by Dr. Nathan Redman, successfully developed and scripted version 1.1 of a ResNet-style architecture neural network utilizing the PyTorch framework. This version includes advanced features such as cross-entropy loss and a resource management CUDA module, ensuring efficient processing and management of computing resources.

3. Image Preprocessing Development:
Dr. Nathan Redman has begun initial scripting and development of the image preprocessing module, crucial for preparing raw histology slide data for neural network analysis. Utilizing Python Image Library and OpenCV, this module ensures that the raw image data are appropriately formatted to be optimized for input into the neural network.

4. Collaboration Changes:
Dr. Liam Neiswanger-Broughton of Washington State University had initially agreed to participate as a collaborating histopathologist, but has had to step away from the project. Discussions are currently underway with other potential histopathologists to assist with the manual ground-truth scoring of histological slides, a critical component for training our model.

Challenges and Adjustments
The delayed formal signing of the project contract posed initial administrative challenges; however, the team adapted quickly, ensuring that project milestones remained on track. The unexpected withdrawal of Dr. Broughton-Neiswanger necessitated a search for additional expertise, which is currently being addressed to minimize impact on the project timeline.

Next Steps
For the upcoming quarter, the focus will be on:
1. Finalizing collaborations with new histopathologists.
2. Beginning the collection and preprocessing of histology slides as per the project timeline.
3. Further refining and testing of the neural network model to ensure robustness and accuracy.

View uploaded report

Wet lab tests and farm trials that test effects of any parameter (e.g., breeding programs, supplements, probiotics, antinutritional factors) on distal enteritis are critical for increasing levels of soy protein replacement in fish feed (Booman, 2018). Professional grading of slides costs tens of thousands of dollars and may vary in quality and time requirements. This approach makes the grading process instantaneous, reliable above 95%, and cost-free, removing a hurdle to increase soy utilization in aquaculture.