Technical Approach

I chose skin lesion detection as the purpose of my AI because:

• Skin cancer is a significant issue: around 1 in 4 Americans die from skin cancer

• According to the skin cancer foundation, early skin cancer detection saves lives.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

To train the AI to detect skin lesions, I needed a pre-trained model to fine-tune for skin lesion detection and a dataset of images to fine-tune the model with. I used the Resnet-18 pre-trained model as my model, the ISIC 2020 skin lesion competition as my dataset, and the Matlab Deep Network Designer to fine-tune the pre-trained model.

​

I imported the Resnet-18 model and the ISIC 2020 Skin Lesions Dataset in deep network designer, set the model's settings to match the dataset, and iteratively trained the Resnet-18 on the dataset while changing around the training settings to get the highest accuracy possible.

 

During the last two weeks of the internship, I drafted a formal report on the Internship and made a presentation, which I would go on to present to my advisor, Venkat Krovi, my collaborator, Aditya Krovi, and a bunch of students at the ICAR Labs.

​

Going forward, I will edit and revise the paper, and I plan to publish the AI on GitHub in my repo. The fine-tuned model has the possible potential to, if it is developed to a stage of 90% or more accuracy on the skin lesion dataset, be uploaded to the cloud to be implemented in a skin lesion detection app.

​

​