COVID-19 Scan Classification Experiments

Repository of machine learning experiments for binary image classification (COVID vs Non‑COVID) using transfer learning, plus a few supporting exploration scripts (PCA + GMM segmentation).

Note: This repo contains research/experiment scripts (some written in a notebook-like style) and includes hard-coded local paths to datasets/checkpoints that are not part of the repository. The README below shows how to adapt the scripts to your environment.

---

What I built

• Binary classifier (COVID vs Non‑COVID) using transfer learning with VGG16 (TensorFlow/Keras).
• Feature exploration using PCA on grayscale image vectors (and notes toward PCA on deep features).
• Image segmentation toy example using a Gaussian Mixture Model (GMM) on pixel colors.
• Checkpoint loading experiments (e.g., loading pretrained/previously trained weights).

---

Tech stack

• Python
• TensorFlow / Keras
• NumPy, Pandas
• scikit-learn
• OpenCV
• Matplotlib (and optionally Seaborn for plots)

---

Repository structure

• transfer_learning_vgg.py: VGG16 transfer learning workflow (data split into train/valid/test, training + evaluation, confusion matrix/ROC plotting).
• load_greyscale_inception.py: Loads InceptionV3 (without top) and attempts to restore weights from a checkpoint directory.
• image_preprocessing_pca.py: Grayscale preprocessing + PCA variance plot + additional PCA exploration snippets.
• gmm.py: Simple GMM-based image segmentation demo.
• test.py: TensorFlow v1-style checkpoint restore snippet.

---

Dataset expectations

The scripts expect an image dataset organized like:

<DATA_ROOT>/
  train/
    Covid/
    Non-Covid/
  valid/
    Covid/
    Non-Covid/
  test/
    Covid/
    Non-Covid/

Some scripts also start from a “flat” folder of images and then move a random sample into train/valid/test. If you don’t want files moved, copy the logic and replace shutil.move(...) with shutil.copy(...) (recommended).

---

How to run (local)

1) Create an environment

python -m venv .venv
# Windows (PowerShell)
.\.venv\Scripts\Activate.ps1
python -m pip install --upgrade pip

2) Install dependencies

There isn’t a pinned requirements.txt in this repo. A practical starting point is:

pip install tensorflow numpy pandas scikit-learn opencv-python matplotlib seaborn

If you run into version issues, match TensorFlow to your Python version (common on Windows).

3) Update paths inside scripts

Several files contain absolute paths like C:\\Users\\itsios\\Desktop\\dissertation\\....

Search/replace these with your local locations:

• Dataset root (images)
• Checkpoint directory / checkpoint file

4) Run an experiment

Example:

python transfer_learning_vgg.py

---

Results

The accompanying MSc thesis evaluated three VGG16‑based transfer learning strategies (without fine‑tuning) for binary COVID vs Non‑COVID CT classification:

• Method 1: Train only the final dense classification layer on top of a frozen VGG16 backbone.
• Method 2: Transfer learning from the last five convolutional layers of VGG16, adding new pooling + dense layers trained on COVID CT scans.
• Method 3: A refined architecture building on Method 2, with a carefully tuned learning rate and capacity to further improve generalization.

On the held‑out test set, the models achieved:

Metric	Method 1 (last layer only)	Method 2 (last 5 conv blocks)	Method 3 (refined architecture)
Accuracy	0.94	0.97	0.99
Precision	0.96	0.96	1.00
Recall	0.92	0.98	0.98
F1‑score	0.94	0.97	0.989
AUC	0.94	0.97	0.99

Summary:

• Progressive model design: each successive method improves or matches the previous one on all key metrics.
• Best model (Method 3) delivers 99% accuracy and AUC 0.99 with perfect precision (1.00), indicating very few false positives.
• Robust evaluation: results are supported by confusion matrices, ROC curves and cross‑validation as documented in the thesis.

---

Highlights

• Transfer learning approach to reduce training time and data requirements.
• End-to-end experimentation: preprocessing → training → evaluation metrics → visualization.
• Model introspection ideas: extracting intermediate layer outputs (e.g., VGG fc2) to study representations.

---

Limitations / Notes

• Scripts are experiment-oriented and may include notebook-only directives (e.g., %matplotlib inline) and deprecated APIs.
• The code assumes external assets (datasets/checkpoints) not included in the repository.
• This project is for educational/research purposes and is not a clinical tool.

---

License

If you plan to publish this on GitHub, add a license you’re comfortable with (e.g., MIT) and ensure the dataset (if any) is licensed for redistribution.