Summary:
AI Tumour Detection
The goal of this project was to create a machine learning model that can be trained on patient data while keeping that data private. This was done using a federated learning model which consists of multiple "local" clients that can be sent out to hospitals or other medical centres, and a single central server that is used to aggregate the individual models once their training is complete.
The aggregation server was able to classify tumours with a max accuracy of ~60% with 4 clients being used.
Although this accuracy seems to be quite low in this context, this is mostly due to the small size of the dataset that was available, as well as the lack of consistency in the data that was fed into each of the clients.
The datasets included multiple different types of tumours as well as scans from different orientations such as top-down, lateral slices, etc. If the data being fed into the clients was more consistent, a much higher accuracy would be achieved.
In a real-world implementation, a technician at each medical facility would only feed in images of the same orientation such as top-down scans. Additionally, installing clients in as many medical facilities as possible, would drastically improve the classification accuracy due to the larger dataset that could be collected.
Technical Explanation:
