A novel unsupervised segmentation approach for brain computed tomography employing hyperparameter optimization

This work proposes a methodology for segmenting structures and brain tissues in computed tomography scans using unsupervised deep learning. The methodology involves extracting features from the CT scans and applying similarity and continuity constraints, creating segmentation maps of intracranial structures and observable tissues. This approach can assist experts in diagnosis by identifying specific regions with anomalies. The method is applied to a database of real scans and uses a spatial continuity evaluation function directly related to the desired quantity of structures. Results demonstrate that the proposed unsupervised methodology achieves segmentation of the desired number of labels and allows for a reduction in effort compared to supervised segmentation models.