Dissertation: Model Repo

Model architecture, code and weights for MSc Dissertation Project: "Mobile-Optimised Deep Learning Framework for Interpretable Detection of Fungal Keratitis via Lateral Flow Device Imagery".

Important Info

The original dataset used in this dissertation involved private medical data and has been stripped to preserve privacy. Users can insert their own data into src/data/ instead.

What's in this repository?

This repository contains the full codebase for the training and evaluation of deep learning models developed for the dissertation project. It includes scripts for data processing, model training, and performance benchmarking, along with the saved results and model weights.

Organisation

• experiments/ - Contains additional visualisations and tests run to infer data saved in runs.
• grad-cam/ - Stores results of Grad-CAM runs of models to see their region of focus on images.
• models/ - Model architecture code for fine-tuning and other methodologies.
• results/ - Stored metrics of test-set evaluation - AUC-PR, AUC-ROC, F1.
• scripts/ - Scripts used for headless, cloud training.
• util/ - Utility functions needed for runs.
• weights/ - Saved model weights of best runs for each methodology per architecture.

Setup

Create a fresh python virtual environment and run install.py to install all the required dependencies. Alternatively, you can manually install the required dependencies using the requirements.txt file.

Runs

Python bench supports training, validation and evaluation for different methodologies: Conventional classification (fine-tuning and training from scratch), Siamese Few-Shot training, Knowledge Distillation. The training regime and models used are selected based on command-line arguments. Here are a few examples of how to run the bench (assuming the data is already prepared):

1. To run a classification cross-validation with no additional mechanisms (baseline):

python runner.py --task_type classification_crossval --models mobilenetv3 efficientnet shufflenet --mode train --epochs 60 --lr 1e-4 --batch 32 --workers 8 --patience 10 --label cross_val_baseline

2. To run a traditional classification to save model:

python runner.py --task_type classification --models efficientnet --mode train --epochs 60 --lr 1e-4 --batch 32 --workers 8 --patience 10 --roi --weighted_sampling --label efficientnet_roi_weighted_sampling

3. To run an evaluation on a saved model:

python runner.py --task_type classification --models efficientnet --mode evaluate --batch 32 --workers 8 --patience 10 --roi --weighted_sampling --label efficientnet_roi_weighted_sampling --file efficientnet_fold1_2025-06-16_15-05-49.pth

Since the test bench switches the type of training, regime used and models trained based on the command arguments passed, please refer to the code and scripts/next.sh for more parameters and running commands.

📊 Results

Held-out Test Set Performance

• EfficientNet-B0

  • Accuracy: 93.3%
  • Precision: 1.00
  • Recall: 0.667
  • F1: 0.80
  • Validation Loss: 0.47 (threshold = 0.56)

• ShuffleNetV2

  • Accuracy: 86.7%
  • Precision: 1.00
  • Recall: 0.333
  • F1: 0.50
  • Validation Loss: 0.51 (threshold = 0.47)

---

AUC Metrics (threshold-independent)

• EfficientNet-B0: PR-AUC 0.880, ROC-AUC 0.940
• ShuffleNetV2: PR-AUC 0.650, ROC-AUC 0.760

---

Mobile Inference (iPhone 11 Benchmarks)

• All models ran fully offline with no network calls.
• ShuffleNetV2 showed the lowest latency and memory usage (< 500 MB).
• Ensemble inference was slower but still within device constraints.

---

Interpretability

• Grad-CAM visualisations confirmed EfficientNet focused consistently on the diagnostic lines of the LFDs.
• ShuffleNet produced less precise attention maps.