AlexNet for Skin Lesion Classification

This repository contains the code for the project Fine-Tuning and Boosting AlexNet for Skin Lesion Classification. The primary objective is to classify skin lesions using an AlexNet-based deep learning architecture, fine-tuned with physical data augmentation techniques to enhance performance.

link to the competition : https://www.kaggle.com/competitions/ima205-challenge-2024/leaderboard

Detailed report is in : IMA_205_Project_report_AlexNet_for_Skin_lesion_classification

Overview

Skin cancer, particularly malignant melanoma, is a significant health concern, accounting for the majority of skin cancer-related deaths. Early detection is crucial to improve patient outcomes. This project leverages a deep learning model, AlexNet, to classify different types of skin lesions from the ISIC dataset. The project involved extensive data preprocessing, augmentation, and fine-tuning of a pretrained AlexNet model to optimize classification performance.

Project Goals

• Utilize the ISIC dataset to train and fine-tune a neural network for skin lesion classification.
• Apply various data augmentation techniques to improve model generalization.
• Achieve efficient and accurate classification while minimizing computational resources.

Dataset

• Training Set: 18,998 images
• Test Set: 6,333 images
• Classes: Melanoma, Melanocytic nevus, Basal cell carcinoma, Actinic keratosis, Benign keratosis, Dermatofibroma, Vascular lesion, Squamous cell carcinoma

Key Techniques

Data Preprocessing

• Hair Removal: Morphological closing and hysteresis thresholding to remove occlusions.
• Data Augmentation: Various transformations such as flipping, rotation, brightness adjustment, noise addition, and more to prevent overfitting and enhance model performance.

Model Architecture

We used a modified AlexNet architecture for classification with 8 output classes. The architecture includes:

• Convolutional layers with ReLU activation
• Max Pooling
• Fully connected layers
• Softmax output for classification

Evaluation

The primary metric used for evaluation was Weighted Categorization Accuracy (WA), which considers class imbalance in the dataset.

Results

• Initial model accuracy without augmentation: < 50%
• Final model accuracy after augmentation: 69%
• Ranking: 13th out of 80 in a school Kaggle competition

Challenges

• Integration of metadata with image data.
• Time constraints and limited access to resources.
• Handling dataset instability due to the use of patches.

Future Work

• Experimenting with alternative architectures like multi-class SVMs.
• Incorporating more metadata for further performance improvements.

Requirements

• Python 3.12
• Libraries:
  • torch
  • numpy
  • pandas
  • matplotlib
  • scikit-learn
  • albumentations for data augmentation

Install dependencies using:

pip install -r requirements.txt