Skin Cancer Classification

The goal of this project is to classify images into Benign or Malignant category.

Gemini Agent Based Classification

streamlit run agent.py

Gemini-based AI agent to assess skin lesions using the ABCDE criteria — Asymmetry, Border irregularity, Color variation, Diameter, and Evolution/Elevation. The agent not only evaluates each criterion but also provides clear, patient-friendly explanations, making complex medical insights accessible and understandable.

Dataset Exploration

First look at some images from the dataset

One can see three main challenges in our dataset:

• Hair occlusion: Visible hair strands obscure important lesion features

• Artificial borders: Black circular borders around lesions create non-biological patterns

• Imbalanced classes: Uneven distribution between benign and malignant samples may bias the model

  Preprocessing Techniques

  Hair Removal

  To combat this issue created an adaptive DullRazor algorithm combined with few other image processing techniques to effectively remove hair while preserving crucial lesion details.

Model Architecture and Training Techniques

I implemented EfficientNet-B0 with Focal Loss to address the class imbalance problem. The training approach uses discriminative learning rates - lower rates for pre-trained backbone layers and higher rates for the classifier layers.

For optimization, I employed AdamW with weight decay and OneCycleLR scheduling, which gradually increases learning rate during the first 30% of training before decreasing it, allowing for efficient convergence while maintaining stability.

Data augmentation techniques include random horizontal and vertical flips (20% probability each), slight rotations (±10°), random translations (up to 10%), scale variations (90-110%), and color jittering to improve model robustness and prevent overfitting. Images were resized to 232px and center-cropped to 224px before normalization using ImageNet statistics.

Results

I was able to achieve 94.6% accuracy in Test dataset

Class	Precision	Recall
Benign	0.936	0.957
Malignant	0.956	0.935

Overall Accuracy: 0.946

Usage

The preprocessed skin cancer dataset (with hair removal and border smoothening applied) is available for download:

Download Preprocessed Dataset

To train the skin cancer classification model on your dataset:

python skin_classifier/train_model.py --root_dir /path/to/your/processed/dataset

To evaluate the model on test data:

python skin_classifier/test_model.py --model_path /path/to/your/model.pth --test_dir /path/to/test/data

Please refer to the code for more customizable arguments and advanced configuration options.