SSEGEP: Small Segment Emphasized Performance evaluation metric for medical image segmentation

Abstract

Automatic image segmentation is a critical component of medical image analysis, and hence quantifying
segmentation performance is crucial. Challenges in medical image segmentation are mainly due to spatial
variations of regions to be segmented and an imbalance in distribution of the classes. Commonly used
metrics treat all detected pixels indiscriminately. However, pixels in smaller segments must be treated differently from pixels in larger segments, as smaller ones aid in early treatment of associated disease and are
also easier to miss. To address this, we propose a novel evaluation metric for segmentation performance,
emphasizing smaller segments, by assigning higher weightage to smaller segment pixels. “Small SEGment
Emphasized Performance” (SSEGEP) metric, based on weighted false positives, was proposed and evaluated using synthetic images and 4 publicly available clinical datasets of eye (fundus imaging, n = 33),
breast (mammogram, n = 108), liver (CT, n = 131), and pancreas cancer (CT, n = 107), where n refers
to the number of images from the dataset that were used in the study. Mean opinion score (MOS) was
calculated from the scores (5-scale scores) of 33-fundus-image segmentation as assigned by 15 researchers
(2–5 years of experience in image analysis). Statistical analysis was performed for the other datasets to
quantify the relevance of the proposed approach. Across 33 fundus images, where the largest exudate is
1.41% and the smallest is 0.0002% of the image, the proposed metric is 30% closer to MOS, as compared
to Dice Similarity Coefficient (DSC). Statistical significance testing resulted in a p value of order 10−18 that
shows the significance of SSEGEP for hepatic tumors compared to DSC. The proposed metric is found to
perform better for the images having multiple segments of a single class.