As the name entails, this research project talks about how AI can be used to detect cancer specifically prostate cancer. To achieve this, The type of model used for this project was a convolution neural network or CNN. As shown in Figure below, a convolutional neural network takes in input images, and by a string of different techniques it converts them to numbers for the neural network to process. CNNs are used almost entirely for image classification. As shown in the Figure there is also a step called convolution in which it filters the image and extracts features from it. Like most neural networks there are then various different hidden layer techniques that can impact the performance of the model. 

The specific model we used for our project was Resnet18. Resnet is a very popular set of neural network models that are known for their efficiency. Resnet 18 specifically uses 18 layers. It was shown to be very efficient and also relatively computationally heavy. A figure of the network's architecture is also shown below.

Despite a few challenges found with a small dataset, we were able to train the model to around 85% accuracy. This is very good and shows how effective this technology can be in the future. We used data from medical decathlon. The data was greyscale and overall a pretty small dataset. We converted the data from grayscale to color in order to use the 3 input layer that was required for resnet 18. A sample batch size is shown below of 20. Different from the pi-rads score, this data set can have images accompanied by either a 0 or 1. Meaning the images were only classified as benign or malignant. The batch size below has both benign and malignant prostate cancer scans in them but to the untrained eye they are hard to tell the difference. 

Attribution maps show what attributes of the image activate the model and cause classification. As we can see in one of the results below, There is activation in the middle towards the bottom. This means that something in that spot of the original image is activating the network and helping it classify. The most important reason for this part is to show not only that our model works, but it works for the right reasons. Since this technology is not being used to replace doctors but rather help them, it is important to have regions in which prostate cancer is thought to be present in order to speed up the process of the doctor who is reviewing it. The attribution maps also had slight post processing done to them to improve the results.

We used a few different types of attribution maps. Saliency, noise tunnel applied to saliency, and Input X Grad. Saliency maps highlight the parts of the input data that are most important for the network's decision-making process. Noise tunnel applies a map to the saliency map and highlights the location most susceptible to change. Meaning the location of the image that will change its classification if the thing being highlighted is present or not. Input X Grad is another attribution map tool. They provide different results by calculating the gradient of the output of the network with respect to the input data and then taking the absolute value of the gradient. This creates different data from the first two attribution maps and will ensure that the model is accurately understanding the image and making correct classifications.

Our future plans for this AI Prostate Clinic plans would be to continuously to improve its detection system. The detection system will become more accurate as the training data improves. Our team is striving to reach as close to 100% accuracy as possible for our trained data. Another thing that will improve our clinic is if we get access to prostate data from a Cancer Center we reached out to early this semester. If we can get access to new prostate cancer data then we will be able to make a new detection system with it or combine it with our current data and retrain the model. I will keep this page updated as we improve our network and the code snippets will soon be available on the snippet page.