[PROJECT] Web-based skin cancer (melanoma) detection

[shkimmie-umb]

Overview

• Melanoma is the most severe type of skin cancer, and an automated melanoma diagnosis method helps patients find and cure melanoma in advance, reducing significant chances of mortality.
• Melanoma exists in various shapes and colors, and deep learning-based melanoma detectors can tackle the deviations of melanoma shapes when detecting them.

Challenge

• Deep learning models require lots of data to detect the desired shapes correctly. Still, the number of available melanoma datasets is limited, as skin cancer data is managed and stored in different forms and different methods of metadata, so it is not easy to combine them.
• Existing deep learning-based skin cancer detectors use only limited data to learn the detector, and performance tests have been conducted under inconsistent conditions, making it difficult to objectively determine which one is more effective even if many detectors have been introduced so far.
• Therefore, developing a high-performance skin cancer detector that can be used in a medical environment has been difficult.

Our approach

• To the best of our knowledge, our skin cancer detector generated 1298 experiments on all public skin cancer datasets and 28 different deep learning models before creating a real-world objective highest performance detector.
• By deploying the highest performance detector to the web, we created a web-based detector that is not restricted by the medical staff's medical environment.
• We considered a model that trains new skin cancer data to our detection model in real time and enables continuous performance improvement. Existing top-performance detectors ResNet and DenseNet are too large to meet these requirements. Therefore, we created a model based on the MeshNet model and compared its performance.
• As a result, a Mela-D model with a model size reduced by more than 20 times while showing similar performance was obtained.

Current challenges

• We're constantly improving our classifier to perform comparably to 1st ranked melanoma classifier from SIIM-ISIC Melanoma Classification. There is room for improvement in terms of performance by trying diverse methods such as:
  • Image resizing
  • Augmentation
  • Hyperparameters
• Improving Mela-D classifier further

Web-based Melanoma Detection

• ArXiv link

Standalone Melanoma web application

Melanoma web application

Original Repository

Melanoma Repo