An advanced Deep Learning system designed for the automated classification of skin lesions into Benign and Malignant categories, as well as multi-class diagnosis. This project leverages state-of-the-art Convolutional Neural Networks (CNNs) to provide diagnostic support in dermatology.
This implementation strictly follows professional deep learning standards for medical imaging:
- Architectures Explored: * Custom CNN: Baseline model for performance benchmarking.
- Transfer Learning: Implementation of ResNet-101, DenseNet-101, and EfficientNet-B3 for feature extraction.
- Data Pipeline: * Sophisticated Augmentations: Flips, rotations, color jitter, and random resized cropping to improve generalization.
- Handling Imbalance: Optimized loss functions and weighted sampling for medical accuracy.
- Optimization: Adam optimizer combined with advanced learning rate scheduling.
- Medical Accuracy: Achieved high F1-Score and ROC-AUC by fine-tuning backbones on medical datasets.
- Error Analysis Module: Developed a custom visualization script to analyze "Incorrectly Predicted" samples, allowing for qualitative clinical auditing.
- Inference Optimization: Conducted experiments to balance model parameters (M) vs. inference speed (img/s).
- Compute: Trained using NVIDIA T4 GPU acceleration.
- Platform: Developed and tested on Google Colab for cloud-based scalability.
- Metrics: Macro-Precision, Macro-Recall, and Macro-F1 Score.
Model/: Core Jupyter Notebook with full training logs.Results/: Evaluation metrics and performance comparison tables.
Developed as part of the Deep Learning (WDL) Specialization. Copyright © 2026 Mahmoud Souliman. All rights reserved.