Back to Homepage

Deploying AI within Ethological Research for a Neuroscience Laboratory

A seminal collaboration with Sahlgrenska University Hospital, my Bachelor's thesis addressed the critical imperative to automate the labour-intensive and subjective manual decoding of animal behaviour. I engineered and deployed a comprehensive computer vision framework utilising Convolutional Neural Networks to automate the tracking of mice in ethological experiments. My primary contribution encompassed the end-to-end architecture, execution, and scoping of the DeepLabCut implementation, delivering a scalable solution that significantly enhances the precision and throughput of neuroscientific data acquisition.

Project III

password: "mouse"

< 5px Error Rate
Objective Scoring
~22ms Inference
0 Markers Used

Technical Stack

Python DeepLabCut TensorFlow OpenCV Streamlit MoviePy Google Colab NVIDIA A100

Bespoke CNN Architecture

Engineered and fine-tuned Convolutional Neural Networks (CNNs) with custom architectures for markerless pose estimation, specifically optimised for identifying and classifying distinct animal behaviours with high accuracy from video streams.

Autonomous Behavioural Decoding

Developed a robust Python-based application that automates the analysis of video data to decode complex animal behaviours, significantly reducing the need for time-consuming and subjective human intervention.

Advanced Multi-stage Pre-processing Pipeline

Designed and implemented a multi-stage pre-processing pipeline, including temporal segmentation, spatial isolation, and image quality improvement. This involved developing a custom web-based application (using Streamlit) for timecode and coordinate annotation, and Python scripts (utilising libraries like OpenCV and MoviePy) to standardise raw video data for optimal AI inference.

Robust AI Inference and Data Interpretation

Leveraged the DeepLabCut framework with pre-trained SuperAnimal models (specifically TopViewMouse) for markerless pose estimation. The system utilised high-performance computing (NVIDIA A100 GPUs on Google Colab Pro) for efficient inference. A dedicated Python post-processing script was developed to interpret the raw HDF5 output, translating it into actionable behavioural metrics (e.g., locomotion, chamber transitions) that align with expert human coding protocols, whilst also addressing inherent variabilities in human observation.

Empirical Neuroscience Assistance

Provided practical and scalable technological solutions directly to Sahlgrenska University Hospital’s neuroscience laboratory, enhancing their data collection capabilities, facilitating novel insights into neurological conditions, and contributing to more objective and high-throughput research.