Overview / Usage

This project is a deep learning model that aims to visually diagnose melanoma, the deadliest form of skin cancer. In particular, the model distinguishes between this malignant skin tumor from two types of benign lesions, nevi and seborrheic keratoses.

The problem being solved here is trying to come up with a noninvasive, easy to use system that can be used by actual dermatologist in their work. It's also meant to reduce miss classification of lesions and reduce diagnosis time. Therefore, this project finds it's use in Healthcare, especially in dermatology.

Methodology / Approach

Methodology and Architecture:

1. Dataset. The data and objective are pulled from the 2017 ISIC Challenge on Skin Lesion Analysis Towards Melanoma Detection. As part of the challenge, participants were tasked to design an algorithm to diagnose skin lesion images as one of three different skin diseases (melanoma, nevus, or seborrheic keratosis).

2. Architecture. The model has the DenseNet201 as it's base pre-trained network but tweaked to fit the three classes of the dataset. This pre-trained network was chosen since it was of a smaller size for packaging it to a mobile app. Testing has not yet been done, but through quantization, it should be easier.

3. Back-end Framework. PyTorch

4. Loss and Accuracy. With just 10 epochs, the model was able to achieve an test loss of 1.207794 and an accuracy of 46%, which was 160 images correctly classified over the total 347 images in the testing set. And since it's still under development, it should be able to improve with time.

​

Appoach:

1. The dataset was already split into training, validation, and testing folders each containing the three classes of lesions.

2. Only two forms of transforms were done of the images, which were, center cropping to 120, and transformation to tensors.

3. The DenseNet201 pre-trained model was loaded and classifier tweaked to fit the classes in the dataset.

4. The Cross Entropy Loss was used since it's a classification task, together with Stochastic Gradient Descent as the optimizer.

5. All training was done using PyTorch and on Intel CPU.

Technologies Used

Hardware:

1. Intel® Core™ i5-3320M CPU @ 2.60GHz × 4 Processor
2. Intel® Ivybridge Mobile Graphics

Software and Libraries:

1. Ubuntu 19.10 64 bit
2. Python 3.7.5
3. Powered by PyTorch
4. Intel DevCloud

Repository

https://github.com/martinoywa/dermatologist-ai