Chick Embryo Stage Classifier

ISEF Category: Animal Sciences

Subcategory: Development  ·  Difficulty: Intermediate  ·  Setup: Home Setup  ·  Time: 1 to 2 Months

The Hook

A chick embryo changes fast, but a candling image can still look like a blur of shadows and veins. If you can teach a model to read those patterns, you can turn a basic flashlight check into a stage tracker. That gives you a real machine learning project with a biology question behind it.

What Is It?

Candling means shining light through an egg so you can see what is happening inside without opening it. In this project, you use those images to predict the embryo's developmental stage, which is the point on the growth timeline.

Think of it like sorting trees by rings, except the clues are egg shape, shadow patterns, blood vessels, and how the dark area fills the shell. A classifier is the model that learns those visual cues and assigns each image to a stage label. Public datasets let you test the idea before you ever touch an incubator.

Why This Is a Good Topic

This topic works well because you can test the same image set with different models, preprocessing choices, and label schemes. It connects to hatchery monitoring, embryo research, and computer vision in a way you can explain clearly. You can learn data cleaning, training, validation, and error analysis without starting with a wet lab.

Research Questions

• How does image preprocessing affect stage classification accuracy on chick embryo candling images?
• What is the effect of using cropped embryo regions instead of full egg images on model performance?
• Does a convolutional neural network outperform a classical feature-based model on the same candling dataset?
• To what extent does class imbalance change recall for early developmental stages?
• Which lighting-normalization method reduces prediction error across different candling datasets?
• How does transfer learning affect accuracy when training on a small public dataset?

Basic Materials

• Laptop or desktop computer with at least 8 GB RAM.
• Public chick embryo candling-image dataset.
• Python installed with Jupyter Notebook.
• Spreadsheet software for labels and error logs.
• External storage or cloud drive for backups.
• Optional low-cost egg incubator kit, if you add an ethics-approved hands-on extension.

Advanced Materials

• Research-grade incubator with temperature and humidity logging.
• Fixed LED candling light and camera mount.
• High-resolution camera or microscope camera.
• Calibration target or gray card for image correction.
• Image annotation software.
• GPU-capable workstation.
• Institutional ethics approval documents and handling protocol.

Software & Tools

• Python: Trains baseline models, runs experiments, and scores predictions.
• Jupyter Notebook: Keeps code, plots, and notes in one place.
• scikit-learn: Builds baseline classifiers and evaluation metrics.
• TensorFlow: Trains image classifiers on candling datasets.
• ImageJ: Inspects, crops, and measures image regions before modeling.

Experiment Steps

1. Define the stage labels you will predict and decide whether you need a binary, three-class, or multi-class model.
2. Audit the dataset for label quality, class balance, blur, glare, and repeated eggs before you train anything.
3. Choose a baseline model and a deeper image model so you can compare simple features against a CNN.
4. Plan train, validation, and test splits that keep images from the same egg or incubation run together.
5. Select metrics that match the biology question, such as macro F1, recall for early stages, and confusion patterns.
6. Build an error-analysis plan that ties wrong predictions to lighting, occlusion, or stage overlap.

Common Pitfalls

• Mixing images from the same egg across train and test sets, which makes accuracy look better than it is.
• Training on uneven stage counts, which pushes the model toward the most common class.
• Using images with different light levels, which teaches the model glare instead of embryo features.
• Treating blurry or partly blocked candling images as clean labels, which adds noise to the targets.
• Reporting only accuracy, which hides weak performance on early or rare stages.

What Makes This Competitive

A stronger project goes beyond picking the highest score. You compare simple and deep models, test whether cropping or lighting correction helps, and report per-class errors, not just overall accuracy. If you show that one preprocessing choice reliably improves early-stage prediction across datasets, your study starts to answer a real research question. Careful validation and error analysis matter more than a flashy model.

Project Variations

• Use only public datasets and compare a CNN, a random forest, and a support vector machine on the same stage labels.
• Add a preprocessing study that tests whether glare reduction or egg cropping improves stage prediction more than raw images.
• If ethics review allows, collect a small home-incubator dataset and test whether a model trained on public data transfers to your own images.

Learn More

• PubMed: Search review articles on chick embryo staging, candling, and developmental imaging.
• PubMed Central: Read full-text papers on embryo image classification and computer vision methods.
• NCBI Bookshelf: Find open biology chapters on embryonic development and staging.
• MIT OpenCourseWare: Review free machine learning lectures for training, validation, and model selection.
• ImageJ documentation: Learn basic image inspection, cropping, and measurement tools for image preprocessing.