AI-Driven Skin Cancer Detection Application

Project Objective

To develop a skin cancer detection application that can assist dermatologists and individuals in identifying potential skin cancer lesions from images of moles. The goal is to create a reliable and accessible tool for early detection and diagnosis, which can potentially save lives through early intervention.

90% accuracy on the validation set

37 dermatologists in focus group

Partner's Profile

Accuracy and Metrics

The trained skin cancer detection model achieved an accuracy from 53% to over 90% on the validation set. It demonstrated high precision and recall for both benign and malignant cases, indicating its effectiveness in diagnosis.

The web-based application received positive feedback from 37 dermatologists and users alike. Its user-friendly interface and accurate predictions contributed to its adoption and usage for preliminary assessments of skin lesions.

Ongoing Improvement

By providing an accessible tool for early skin cancer detection, the application has the potential to aid in early intervention and improve the chances of successful treatment.

The model continues to be updated with new data to enhance its accuracy and keep up with evolving medical knowledge. User feedback and additional features, such as a skin lesion severity score, are being considered for future application iterations.

Solution

The combination of Scikit-learn, Keras, and the ResNet50 model was used to develop an effective skin cancer detection application. Our team built an AI-powered web application that aligns to leverage machine learning and deep learning for early disease detection and highlights the importance of ethical considerations and continuous improvement in healthcare applications.

Web-Based Application

Creation of a user-friendly web-based interface using Django. Integration of the trained model into the application. Ability for users to upload mole images for classification. Displaying the classification results with an indication of the likelihood of malignancy.

Feature Extraction with ResNet50

Utilization of the ResNet50 model pre-trained on a large dataset to extract high-level features from moles images. Removal of the top classification layers to retain the feature extraction capability.

Custom Classifier

Development of a custom neural network classifier using Keras. The classifier takes ResNet50 features as input and outputs skin cancer classifications (benign or malignant). Training the classifier on the preprocessed dataset.

Model Evaluation

Evaluation of the model's performance using metrics such as accuracy, precision, recall, and F1-score on the validation set. Hyperparameter tuning to optimize the model's architecture and training parameters.

Data Collection and Preprocessing

A curated dataset of moles images obtained from various sources. Splitting the dataset into training, validation, and test sets. Preprocess the images through resizing (500x500), format normalization (converting to RGB), and data augmentation to improve model generalization.