Description

This project explores various neural network models to automatically classify between 3 types of brain tumors from MRI scans. This can help provide diagnosis suggestions which can release healthcare workers from spending time reading MRI scans so that they can spend more time on patients. In addition, tumors are detected and segmented from the MRI scans, which can be useful for pinpointing the exact location of the tumor to support treatment planning.

The data contains 3,064 images of 3 brain tumor types from 233 patients. Data cleaning and preprocessing are performed to convert the data into a consistent format, including resizing images, converting data to NumPy array types, applying one-hot-encoding on response variable, performing data normalization, splitting data into train and test sets, oversampling to rebalance data, and augmenting data. For brain tumor classification, baseline FFNN and advanced CNN are implemented. CNN is visualized using activation maps and saliency maps. For tumor detection and segmentation, UNet, LinkNet, and DeeplabV3+ are implemented.