Enhancing Recall in Skin Lesion Classification with Node Pruning

This project investigates the use of node pruning within a stratified k-fold cross-validation framework to improve the recall of CNNs in medical imaging tasks.

Key Features

• Node Pruning Implementation: Reduces model complexity and enhances recall
• ResNet50V2 Base Model: Utilizes a state-of-the-art architecture for binary classification of skin lesions.
• Comprehensive Evaluation: Assesses performance metrics across multiple pruning strategies.

Repository Structure

• notebook/: Contains the Colab notebook for experiments.
• data/: Includes relevant datasets used for training and evaluation.
• figures/: Visualization assets like bar plots and scatter plots.
• README.md: Project overview and usage instructions.

Results Summary

Pruning strategies improved recall while maintaining AUC. Key metrics are summarized below:

Pruning Strategy	Accuracy Mean	AUC Mean	Precision Mean	Recall Mean
Baseline	0.7323	0.8131	0.7798	0.6540
Pruned_All	0.6237	0.8028	0.5950	0.9167
Pruned_Top1	0.7359	0.8119	0.7591	0.6978
Pruned_Top5	0.7191	0.8113	0.6958	0.7987
Pruned_Top10	0.6825	0.8107	0.6397	0.8821
Pruned_Top15	0.6377	0.8061	0.6036	0.9101

Data Acquisition

To reproduce the results, you’ll need to download the dataset from the Kaggle ISIC 2024 Challenge. Follow these steps:

1. Install the Kaggle API:

   pip install kaggle

2. Obtain your Kaggle API key:

   • Go to Kaggle.
   • Navigate to your account settings.
   • Select "Create New API Token" to download the kaggle.json file.

3. Place the kaggle.json file in the appropriate directory:

   mkdir ~/.kaggle
   mv /path/to/kaggle.json ~/.kaggle/
   chmod 600 ~/.kaggle/kaggle.json

4. Download the dataset using the Kaggle API:

   kaggle competitions download -c isic-2024-challenge

5. Extract the dataset:

   unzip <downloaded-dataset>.zip -d data/

How to Run

1. Clone this repository:

   git clone https://github.com/DevDizzle/node-pruning-research.git

2. Navigate to the project directory:

   cd node-pruning-research

3. Install the required dependencies:

   pip install -r requirements.txt

4. Open the notebook:

   • Navigate to notebook/node-pruning-cnn-analysis.ipynb and open it in Jupyter Notebook or Google Colab.

5. Reproduce the experiments:

   • Follow the instructions within the notebook to run the experiments.

Visualizations

The following visualizations are included in the repository:

• Bar Plot: Average recall across pruning strategies
  

• Box Plot: Distribution of recall across strategies
  

• Scatter Plot: Precision vs. Recall
  

Insights

• Improved Recall: Node pruning, especially the Pruned_All and Pruned_Top10 strategies, significantly improved recall compared to the baseline.
• Precision-Recall Trade-off: The trade-off between precision and recall was effectively managed across pruning strategies.

Citation

If you use this repository, please cite:

@misc{nodepruningresearch,
  author       = {Evan R. Parra},
  title        = {Enhancing Recall in Skin Lesion Classification with Node Pruning},
  year         = {2024},
  howpublished = {\url{https://github.com/DevDizzle/node-pruning-research}}
}