Research Fellowship: Applying AI to Oocyte Analysis in IVF Procedures
An end-to-end research project that culminated in a peer-reviewed publication, a custom data-labeling web application, and a proof-of-concept machine learning model to improve objectivity in fertility treatments.
At a Glance
- Situation: The success of In Vitro Fertilisation (IVF) is limited by the subjective, manual assessment of oocyte (egg cell) quality, which leads to inconsistent outcomes and a lack of standardization.
- Task: To investigate how Artificial Intelligence could be applied to this domain. This involved conducting a formal literature review, developing a tool for data collection, and implementing a proof-of-concept AI model for automated analysis.
- Action: I led three initiatives: 1) A systematic literature review following PRISMA guidelines, resulting in a first-author publication. 2) The development of "MMS", a full-stack Python/Flask web app for expert video annotation. 3) The creation of "AI4Fertility", a computer vision pipeline using OpenCV and a PyTorch CNN to classify oocyte maturation stages.
- Result: The project successfully established a complete research lifecycle. The published paper provides a foundational survey for the field, the MMS tool enables crucial data collection, and the AI model serves as a strong proof-of-concept for automating oocyte assessment, demonstrating the potential for AI to bring objectivity to IVF procedures.
Technical Deep Dive
1. Systematic Literature Review & Publication
The project began with a rigorous academic investigation into the current state of AI in reproductive medicine. The goal was to identify the key challenges, common techniques, and significant gaps in the existing research.
- Methodology: I employed a formal systematic review process guided by the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) statement. This involved formulating precise research questions, defining a search strategy with formal queries for Scopus and PubMed databases, and conducting a multi-stage screening process to filter hundreds of articles down to a core set of the most relevant papers.
- Key Findings: The review highlighted several critical challenges in the field, including the lack of large, high-quality public datasets, non-standardized imaging protocols across clinics, and the need for more explainable AI models. We categorized AI applications into four main areas: Segmentation, Feature Extraction, Classification, and Decision Support.
- Outcome: This comprehensive analysis led to my first-author publication, "On the role of artificial intelligence in analysing oocytes during in vitro fertilisation procedures", in the Elsevier journal Artificial Intelligence in Medicine.
2. Web Application for Data Annotation (Project: MMS)
To address the "lack of data" problem identified in the review, I developed a bespoke tool for expert annotation.
- Architecture: I built MMS (Membrane Mask Selector), a full-stack web application. The backend was developed in Python using the Flask web framework, which handled file uploads, data storage, and serving video frames. The frontend was built with standard HTML, CSS, and JavaScript to provide a simple and intuitive interface for medical experts.
- Functionality: The application allows an embryologist to upload a video of an oocyte. The backend processes the video and presents it frame-by-frame. The expert can then navigate through the video and assign a specific maturation stage label (e.g., 'GV', 'MI', 'MII') to each frame. This process generates a structured JSON file containing the image-to-label mappings, creating a machine-learning-ready dataset from raw video.
3. Computer Vision Pipeline (Project: AI4Fertility)
The final stage was to use the annotated data to build a proof-of-concept classification model.
- Image Pre-processing: The first step in the pipeline was to isolate the oocyte. I used OpenCV in Python to perform segmentation. This involved applying a Hough Circle Transform to automatically detect the circular shape of the oocyte in each frame, crop the image to that region of interest, and normalize its size and intensity. This automated pre-processing is crucial for creating consistent inputs for the neural network.
- Model Architecture & Training: I designed and implemented a Convolutional Neural Network (CNN) using the PyTorch framework. The architecture was tailored for this specific image classification task. The model was trained on the dataset generated by the MMS tool to learn the visual features that differentiate the various maturation stages.