Skin Cancer Classification with Modified AlexNet

This repository contains experiments on skin cancer classification using the HAM10000 dermoscopic image dataset. The project explores the use of AlexNet-inspired CNN architectures with different activation functions, comparing the baseline ReLU to a custom activation function proposed in a referenced journal.

Repository Structure

• skincancer (A).ipynb – Baseline AlexNet-inspired CNN implementation using ReLU activation (~75% accuracy).
• hore siscer pt 2 (B).ipynb – Modified CNN using the custom activation function $$y(x) = (x e^x) \cdot \tanh(\text{softplus}(-x))$$ from the referenced journal (~76% accuracy).

Dataset

• HAM10000: 10,015 dermoscopic images of skin lesions, 7 classes.
• Preprocessing: resizing, normalization, augmentation.
• Imbalance handling: SMOTE oversampling applied to balance classes.

Methods

1. Baseline (Notebook A)

   • AlexNet-inspired CNN with ReLU activation.
   • Input size: 64×64.
   • Achieved ~75% accuracy.

2. Modified Model (Notebook B)

   • Custom activation function from journal: $((x e^x)\cdot \tanh(\text{softplus}(-x)))$.
   • Input size: 28×28.
   • Achieved ~76% accuracy.

Evaluation

Both models were evaluated using:

• Accuracy
• Classification Reports (precision, recall, F1-score)
• Confusion Matrices to analyze per-class performance.

Tools & Libraries

• TensorFlow/Keras
• scikit-learn
• pandas
• numpy
• matplotlib
• seaborn

Reference

Rajput, G. (2021). An accurate and noninvasive skin cancer screening based on imaging technique.

Conclusion

• The baseline AlexNet-inspired CNN with ReLU achieved ~75% accuracy.
• The modified CNN with the journal’s proposed activation function slightly improved performance to ~76%.
• Results highlight the role of activation functions in medical image classification tasks.