Lazaro Martull

Skin Cancer Detection Using Deep Learning (ResNet50)

This project builds a deep learning model to classify skin lesions as benign or malignant using transfer learning with ResNet50 and custom convolutional layers.
The goal is to evaluate whether smartphone-quality images can support early detection of melanoma.

---

🧩 My Role

I contributed to:

• Dataset organization & preprocessing
• Transfer-learning model design
• Hyperparameter tuning and training
• Plotting accuracy/loss curves
• Writing model evaluation & interpretation
• Documenting the workflow

---

🛠 Tools & Technologies

• Python, TensorFlow/Keras
• ResNet50 pretrained weights
• Matplotlib for visualization
• ISIC dermatology datasets
• Google Colab / Jupyter Notebook

---

📄 Full Report & Code

This project folder contains:

• 📘 Final PDF Report
• 🧠 training_code.py – clean, portfolio-ready training script

---

🔍 Project Overview

Skin cancer affects millions worldwide, and early detection dramatically improves survival rates.
This project explores whether deep learning can help classify skin lesions using dermatoscopic images.

Using the ISIC 2018 & 2024 datasets, a transfer-learning pipeline was built using:

• ResNet50 as a frozen feature extractor
• Custom VGG-style dense layers
• Data augmentation
• 224×224 RGB image preprocessing

📌 Final performance:

• Accuracy: 86.8%
• Malignant Precision: 71.6%
• Malignant Recall: 36.8%
• Benign Recall: strong and stable

---

📁 Dataset

Images were gathered from:

• ISIC 2018 Challenge Dataset
• ISIC 2024 Archive

Preprocessing steps included:

• Resize to 224×224
• Normalize and batch
• Augment (flip, rotation, zoom)
• 80/20 train-validation split

---

🧪 Preprocessing, Architecture & Training Code

import tensorflow as tf

height, width = 224, 224

train_set = tf.keras.preprocessing.image_dataset_from_directory(
    "content/image_data/xs_train",
    validation_split=0.2,
    subset="training",
    seed=42,
    image_size=(height, width),
    batch_size=10
)

validation_set = tf.keras.preprocessing.image_dataset_from_directory(
    "content/image_data/xs_train",
    validation_split=0.2,
    subset="validation",
    seed=42,
    image_size=(height, width),
    batch_size=4
)

# ------------------ Model Architecture ------------------

base_model = tf.keras.applications.ResNet50(
    include_top=False,
    input_shape=(224, 224, 3),
    weights="imagenet",
    pooling="avg"
)

for layer in base_model.layers:
    layer.trainable = False

from tensorflow.keras.layers import Dense, Flatten

model = tf.keras.Sequential([
    base_model,
    Flatten(),
    Dense(64, activation="relu"),
    Dense(2, activation="softmax")
])

# ------------------ Training Configuration ------------------

from tensorflow.keras.optimizers import AdamW
from tensorflow.keras import callbacks

model.compile(
    optimizer=AdamW(learning_rate=0.0001),
    loss="sparse_categorical_crossentropy",
    metrics=["accuracy"]
)

earlystopping = callbacks.EarlyStopping(
    monitor="val_loss",
    patience=5,
    restore_best_weights=True
)

history = model.fit(
    train_set,
    validation_data=validation_set,
    epochs=100,
    callbacks=[earlystopping]
)

---

📊 Results & Evaluation

Key observations:

• Strong benign classification performance
• Lower malignant recall, likely due to class imbalance
• Training stabilized early because ResNet50 was frozen
• Adding more malignant samples or fine-tuning top ResNet layers could improve recall

---

📈 Accuracy & Loss Curves

The training logs show:

• Validation accuracy plateaued early
• Loss stabilized with little overfitting
• Early stopping correctly restored best weights

(Plots included in the project report.)

---

🚀 Future Improvements

To enhance model performance:

• Fine-tune upper ResNet50 layers
• Add more malignant images
• Test EfficientNet, MobileNet, or ViT models
• Use metadata (age, lesion area) for multimodal learning
• Convert to TensorFlow Lite for mobile deployment

---

📦 Folder Structure

skin_cancer_detection/
│── index.md            # Project write-up (this page)
│── report.pdf          # Final project report
│── training_code.py    # Clean training script (ResNet50 training)

---

🎯 Summary

This project demonstrates the potential of deep learning to support dermatologists and mobile health apps by providing early classification of skin lesions.
While not a substitute for clinical diagnosis, the model shows promising accuracy and provides a foundation for future medical AI research.