MedAL - Skin Cancer Detection System
Deep learning image classification project for detecting malignant skin moles from medical images. The model was built with TensorFlow/Keras using transfer learning, data augmentation, and a recall-first evaluation approach to reduce the risk of missing malignant cases.
Tech Stack
Key Results
We achieved a 97% Recall rate for Malignant cases, ensuring that the vast majority of cancerous moles are flagged for further review by a dermatologist.
| Class | Precision | Recall | F1-Score | Support |
|---|---|---|---|---|
| Benign | 0.96 | 0.64 | 0.77 | 360 |
| Malignant | 0.69 | 0.97 | 0.81 | 300 |
| Overall Accuracy | 79% | 660 |
Model Architectures
We experimented with two distinct architectures to benchmark performance:
1. Transfer Learning (EfficientNetB0)
- Base: Pre-trained on ImageNet (frozen weights).
- Head: Global Average Pooling + Dropout (0.3) + Dense Output.
- Why: EfficientNet provides excellent feature extraction with lower computational cost than ResNet.
2. Custom CNN (Built from scratch)
- Structure: 3 Blocks of
Conv2D+MaxPooling. - Activation: Used GELU (Gaussian Error Linear Units) instead of ReLU for smoother gradient flow.
- Regularization: Heavy Dropout (0.5) to prevent overfitting on the medical dataset.
Engineering Pipeline
Data Preprocessing & Augmentation
Medical datasets are often imbalanced or limited. We implemented a robust tf.data pipeline:
- Resizing: Standardized to
(224, 224). - Augmentation Layer: Random Flips, Rotations (10%), Zooms (10%), and Contrast adjustments.
- Normalization: Rescaling pixel values to
[0,1].
Optimization Strategy
- Loss Function:
BinaryCrossentropy. - Optimizer:
Adam. - Callback:
EarlyStopping(patience=3) to prevent overfitting and restore best weights.
The “Recall-First” Approach
Standard classification uses a decision threshold of 0.5. However, for cancer detection, this is often too conservative.
We analyzed the Precision-Recall Curve and adjusted the decision threshold to 0.37.
- Result: This shift sacrificed some Precision (more false alarms) to significantly boost Recall (catching more cancer), which is the medically correct trade-off.
Tech Stack
- Core: Python, TensorFlow, Keras.
- Data Viz: Matplotlib, Seaborn (Confusion Matrices).
- Metrics: Scikit-Learn (
classification_report,precision_recall_curve).
Usage
# Load the model
model = tf.keras.models.load_model('medal_model.h5')
# Predict on an image
img = load_and_prep_image('mole.jpg')
prob = model.predict(img)
# Apply the medical threshold
is_malignant = prob > 0.37