EffiDerm

An Efficient Deep Learning Model for Skin Cancer Prediction

Accuracy: 93.07%  |  Parameters: ~1.4M  |  Inference: ~47 ms  |  Size: ~5.6 MB

Overview • Features • Dataset • Architecture • Results • Getting Started

---

Overview

Skin cancer is one of the most prevalent cancers globally, and early detection is critical for improving patient outcomes. EffiDerm is a lightweight Convolutional Neural Network (CNN) engineered for automated skin lesion classification from dermoscopic images.

The model classifies 7 distinct types of skin lesions with 93.07% accuracy while maintaining an exceptionally low computational footprint — only ~1.4M parameters and ~5.6 MB in memory. This makes EffiDerm significantly more efficient than conventional architectures such as ResNet-50 (~25.6M params) or VGG16 (~138M params), enabling real-time inference and deployment in resource-constrained healthcare environments.

---

Key Features

Feature	Detail
Lightweight Architecture	~1.4M parameters — 18× smaller than ResNet-50
High Accuracy	93.07% classification accuracy across 7 lesion types
Fast Inference	~47 ms per image — suitable for real-time applications
Mobile-Ready	~5.6 MB memory footprint for edge/mobile deployment
Class Balancing	SMOTE oversampling to address rare lesion class imbalance
Data Augmentation	Rotation, zoom, shift, and flipping for robust generalization
Explainability	Grad-CAM heatmap visualizations for interpretable predictions

---

Dataset

This project uses the HAM10000 (Human Against Machine with 10,000 training images) benchmark dataset.

Property	Details
Total Samples	10,015 dermatoscopic images
Classes	7 skin lesion categories
Original Resolution	600 × 450 px (RGB)
Input Resolution	Resized to 64 × 64 × 3

Supported Lesion Classes

ID	Code	Lesion Type
0	nv	Melanocytic Nevi
1	mel	Melanoma
2	bkl	Benign Keratosis-like Lesions
3	bcc	Basal Cell Carcinoma
4	akiec	Actinic Keratoses
5	vasc	Vascular Lesions
6	df	Dermatofibroma

---

Model Architecture

EffiDerm follows a sequential CNN design optimized for a balance between accuracy and computational efficiency.

Input (64×64×3)
  │
  ├─ Conv2D(32) → MaxPool → BatchNorm
  ├─ Conv2D(64) × 2 → MaxPool → BatchNorm
  ├─ Conv2D(128) × 2 → MaxPool → BatchNorm
  ├─ Conv2D(256) × 2 → MaxPool
  │
  ├─ Flatten → Dropout(0.2)
  ├─ Dense(256) → BatchNorm
  ├─ Dense(128) → BatchNorm
  ├─ Dense(64) → BatchNorm
  ├─ Dense(32) → BatchNorm (L1L2 Regularization)
  │
  └─ Dense(7, softmax) → Output

---

Methodology

1. Preprocessing

• Resized all images to 64 × 64 × 3
• Normalized pixel values to the range [0, 1]
• Applied one-hot encoding to target labels
• Used SMOTE (Synthetic Minority Oversampling Technique) to balance underrepresented classes

2. Data Augmentation

Technique	Configuration
Rotation	±10°
Zoom	±10%
Width/Height Shift	±10%
Horizontal/Vertical Flip	Enabled

3. Training Configuration

Parameter	Value
Optimizer	Adamax (lr = 0.001)
Loss Function	Categorical Crossentropy
Epochs	50
Batch Size	32
LR Scheduler	ReduceLROnPlateau (patience=2, factor=0.5)
Train/Test Split	75% / 25%

4. Evaluation Metrics

• Accuracy, Precision, Recall, F1-Score
• Confusion Matrix
• Training & Validation Loss/Accuracy Curves
• Grad-CAM Heatmap Visualizations

---

Results

Metric	Value
Classification Accuracy	93.07%
Inference Time	~47 ms/image
Total Parameters	~1.4M
Model Size	~5.6 MB

---

Tech Stack

Component	Technology
Language	Python 3.8
Deep Learning	TensorFlow 2.x, Keras
Data Processing	NumPy, Pandas, OpenCV
Visualization	Matplotlib, Seaborn
Machine Learning	Scikit-learn, Imbalanced-learn
IDE	Visual Studio Code

---

Getting Started

Prerequisites

Python >= 3.8
TensorFlow >= 2.x

Installation

# Clone the repository
git clone https://github.com/agilkannan/EffiDerm.git
cd EffiDerm

# Install dependencies
pip install tensorflow numpy pandas matplotlib seaborn scikit-learn imbalanced-learn opencv-python

Usage

1. Place the HAM10000 dataset CSV in ./dataset/hmnist_28_28_RGB.csv
2. Open and run main.ipynb sequentially
3. The trained model will be saved as skin_cancer_model.h5

---

Applications

• Clinical Decision Support — Assist dermatologists with automated lesion classification
• Mobile Screening — Lightweight enough for on-device skin cancer screening apps
• Telemedicine — Enable remote dermatological assessment in underserved areas
• Medical Education — Interactive learning tool for dermatology students and residents

---

Limitations

• The model is trained exclusively on dermoscopic images; performance on smartphone-captured images has not been validated
• Classification accuracy is dependent on the quality and diversity of the training dataset
• Clinical deployment requires further validation through prospective studies and regulatory approval

---

Future Work

• Enhanced explainability through advanced Grad-CAM and attention-based visualizations
• Model optimization for mobile and edge deployment (TFLite, ONNX)
• Extension to smartphone-captured image datasets for broader applicability
• Integration with cloud-based teledermatology platforms for scalable screening

---

Citation

If you use this work in your research, please cite:

@misc{effiderm2026,
  author       = {Agil Kannan},
  title        = {EffiDerm: An Efficient Deep Learning Model for Skin Cancer Prediction},
  year         = {2026},
  publisher    = {GitHub},
  url          = {https://github.com/agilkannan/EffiDerm}
}

---

Author

Agil Kannan

---

If you find this project useful, please consider giving it a ⭐