Fast fully automatic stroke lesions segmentation based on Parzen estimation and µ-law in skull CT images.

Abstract

Stroke is the second leading cause of death worldwide. Those who survive usually experience vision loss, speech, paralysis, or confusion. Agile diagnosis proves to be decisive for patient survival. This paper presents a proposal for rapid stroke segmentation in cranial CT scans, with 1 second, on average, by sample. The segmentation stage uses Parzen window density estimation to classify potentially injured regions in this approach. A proposed adaptation of the µ-law algorithm is applied to enhance damaged areas about healthy brain regions. The results show that the proposed method has the highest mean of accuracy, reaching 99.85%, with a specificity of 99.94%, surpassing classical methods by 16%. On the other hand, the algorithm presented similarity indexes of 93.39% for the Matthews correlation coefficient and DICE of 93.35%. The proposed methodology also compared the results with four approaches that use deep learning; it proved to be equivalent in accuracy, DICE, and specificity, with superior results in sensitivity up to 8% to one of the approaches based on the recent Detectron2 neural network. The results indicate that the proposed method is competitive concerning the approaches already presented in the literature.