No abstract

Abstract. Most feature-based lesion detection and computer-aided diagnosis methods for medical images require representative data of each region of interest (ROI) for parameter selection. Furthermore, the spatial accuracy of the segmentation of the ROIs from the background can significantly affect certain image features extracted from the ROIs. However, requiring spatially accurate manual segmentations of the ROIs to be used as the ground truth is infeasible for large image sets due to the amount of manual work involved. To relax the requirement of spatial accuracy and to enable spatial refinement of coarse manual segmentations to have more representative feature data, a method based on color information and maximally stable extremal regions of lesion likelihoods is presented. The proposed method is quantitatively compared to several segmentation approaches by using a challenging set of retinal images with spatially accurate ground truth of exudates. The experiments show that the proposed method produces good results measured as Dice coefficients between the refined segmentation and ground truth.