Context-Driven Active Contour (CDAC): um novo método de segmentação de imagens médicas baseado em contorno ativo e compreensão contextual

Abstract

New segmentation approaches based on deformable active contour models are frequently developed and tested in different areas of knowledge, to personalize and improve the accuracy in detecting specific regions of interest. Within this scenario, this study aims to create a new segmentation algorithm based on Active Contour Models (ACM), an adaptive method, via context analysis, to define the total energy, called Context-Driven Active Contour (CDAC). This algorithm incorporates the specialist doctor's knowledge when processing markings made by the user on the image, allowing the incorporation of contextual information that an automatic model may not fully capture. This approach seeks to improve segmentation and democratize the functioning of the proposed model. However, the incorporation of specialized knowledge at the beginning of the process aimed to significantly optimize the results, seeking to enhance the algorithm's performance. The performance of the CDAC was evaluated using Computed Tomography (CT) images of the chest and skull, to segment the lungs, in the chest CT images, and areas affected by Cerebral Vascular Accident (CVA), in the CT images, of the skull. Regarding the chest CT dataset, images from healthy individuals with Chronic Obstructive Pulmonary Disease (COPD) and Pulmonary Fibrosis were analyzed, which contributed to a comprehensive and robust assessment of the efficiency of CDAC for the task of lung segmentation. Furthermore, regarding chest CT images, an individual segmentation assessment of the lungs was also considered, since each lung presents distinct structural characteristics and this factor becomes a challenge, requiring the segmentation model to adapt to the specific anatomical and pathological variations of each lung. The main performance metrics presented in the literature were used to evaluate the CDAC. Quantitative performance metrics included Accuracy, Precision, Sensitivity, Specificity, and Matthews Correlation Coefficient. Regarding similarity performance metrics, the Jaccard Index, Dice Coefficient, and Hausdorff Distance were considered. The results of CDAC were compared with those of other state-of-the-art methods, also based on ACM, and demonstrated potential both in the analysis of images of healthy patients and in those that exhibit areas affected by injuries and diseases.