Measuring Spatial Specificity of Multiple Sclerosis Lesion Segmentation using Dice Spectra

Abstract

Multiple Sclerosis (MS), an autoimmune disorder impacting the central nervous system, is increasingly prevalent. In this context, automated tools for MS lesions segmentation could aid healthcare professionals accelerate treatment decisions and monitor disease progression. In particular, detecting small lesions may prove pivotal for early disease detection. However, current evaluation metrics such as the Dice score estimate an algorithm’s overall performance without considering its potential variability in handling lesions of diverse sizes. To address this issue, we propose an original metric, which we call ‘Dice spectrum’ that offers a more accurate estimation of the model’s performance on lesions of varying sizes. This is particularly important for early disease detection, which involves small lesions, for which the model may behave poorly despite high values of the normal Dice score. On the other hand, the Dice spectrum will clearly show the specific model’s performance on such small lesions.

Authors

* External Author