These are the final results of a work developed during the course MO444 (Machine Learning) at the State University of Campinas.
The architecture consists of three main structures: Global Branch, Attention Branch, and Local Branch. The model uses the concept o hard-self-attention to achieve a high-resolution segmentation.
This is the first segment, responsible for classifying the presence of a Brain tumor in the input image. Besides that, from the latent space of this network is extracted one of the inputs for the next segment.
This segment receives two inputs, the original input image and the latent space from the Global Branch. The latent space is used to find de region of max activation, this process is done through a threshold binarization and a max connected component analysis. This region is used to crop the original image.
This is the last segment, responsible for segmenting the affected region. It uses a similar architecture to the Global Branch. The input is the region of interest extracted from the Attention Branch.
The results achieved were acceptable given the conditions. As can be seen from the test samples, the model was able to perform both tasks well.
