Statistical validation metric for accuracy assessment in medical image segmentation

Abstract

Objective Validation of medical image segmentation algorithms is an open question, considering variance of individual pathologies and the related clinical requirements for accuracy. In this paper, we propose a validation metric capable to distinguish between an over and under-segmentation and account for different clinical applications. Materials and methods In this paper, we propose a validation metric representing a tradeoff between sensitivity and specificity. The metric has an advantage of differentiating between an over or under-segmentation which is an important feature for validating large sets of segmentation results, as human inspection is exhausting and time consuming. Although it is oriented to the accuracy measurement it is also closely related to the robustness of a method. Results Features of the metrics are analyzed alongside their medical impact. A set of numerical simulations is performed in order to compare the proposed metric with standardly used discrepancy measures. The metric is illustrated with a clinical case study, presenting accuracy assessment of an algorithm for calvarial tumor segmentation, validated on six patients.

Active shape model unleashed with multi-scale local appearance

“…for hyperrectangle and hyperellipsoid constraints respectively), while a constant β = 0.3 is chosen for our method. The Distance to 'True Boundary' (DtoTB) and Dice Similarity Coefficient (DSC) [16] to the ground truth are used as the metric of segmentation precision. Table 1 shows the experiment results.…”

Distance guided selection of the best base classifier in an ensemble with application to cervigram image segmentation

“…The performance measure comparison on the validation image set of 50 images is shown in Table 1. The mean and standard deviation of p and q (sensitivity and specificity) and dice similarity coefficient (DSC) [11] are computed for each test image. Figure 3 also visually demonstrates such comparison on a test image.…”

Active Volume Models with Probabilistic Object Boundary Prediction Module

“…Therefore the active volume model is fast enough to be used in near real-time 3D reconstruction. Segmentation accuracy is also validated by comparing to gold standard generated by an expert using methods described in [15]. We performed validation on the 3D left ventricle reconstruction based on 82 slices (Figure 2(a)).…”