Skin Cancer AI diagnosis

project interface- benign cell
model structure
the training stats
project interface- malignant cell

INSPIRATION- Cancer will take a physical and mental toll on your body. It is something that no one has to go through, especially when it comes to life-or-death situations. Detection of cancer is the first step to cure, the earlier the detection the better, this is where our AI comes in. This AI is trained to diagnose melanoma (a type of skin cancer) in just seconds with a whopping 92% accuracy.

WHAT IT DOES- All you have to do is upload a clear picture of the lesion (affected area of the body, presumable cancerous) into the program. This image will be passed to the backend convolutional neural network that will output the chances of it being benign or malignant, with 92% confidence. Experienced doctors take approximately 3 weeks to diagnose cancer with just an 86% accuracy.

HOW IT WAS BUILT- The front end of this project was relatively simple, it was built through Tkinter. The backend AI was made through Keras. We built the architecture of the convolutional neural network and trained it for 15 epochs. The training data was 10,000 images which was also augmented to prevent overfitting and to expand the dataset. Training time took around 30 minutes and it was trained in colab, using Tesla T4 GPU. The neural network uses the Adam optimizer and the 'mae' loss function.

CHALLENGES- We wanted to increase the accuracy of the model, as 92% wasn't quite satisfying. But even after training it for longer, the model wasn't crossing that 92% benchmark. We even implemented powerful techniques like transfer learning, but still, the model was just reaching an accuracy of 92%

ACCOMPLISHMENTS- Creating the AI was a tiresome task of experimentation, but after reaching a nice well-performing AI model, we decided that was going to be used in our project. It was a great achievement teaching AI techniques that take a doctor multiple years to master.

WHAT HAVE I LEARNED?- Building the right deep learning is all about experimentation. I personally feel like I've explored every nook and cranny in the world of AI to build the best model. I've also worked on my front end more, by learning tkinter.

WHAT'S NEXT FOR MELANOMA AI DIAGNOSIS- There's always room for improvement. The next goal I have for this project is to learn new techniques to increase accuracy so that it can be deployed to be used in the medical field.