Improving the quality of information systems for diagnosis: an approach to detecting and correcting segmentation errors

Abstract

Research Context: The development of automated Information Systems (IS) capable of segmenting the left ventricle (LV) in cardiac magnetic resonance imaging (MRI) and estimating clinically relevant biomarkers is fundamental to support diagnostic decision-making. Scientific and/or Practical Problem: Many automated IS based on deep learning (DL) are not fully reliable and lack dedicated modules for error detection, which makes them dependent on constant manual inspection and correction. Proposed Solution and/or Analysis: We propose a post-processing method for IS that automatically detects and corrects segmentation errors in LV cardiac MRI produced by DL systems. Detection is performed by combining metrics computed between consecutive time frames of MRI to identify inconsistent segmentations in the temporal dimension; the correction step reconstructs the problematic frame by interpolating nearby segmentations. Related IS Theory: This work is grounded in the perspectives of Information Processing Theory. Research Method: The method was validated on LV segmentations containing both artificially generated and real DL errors, using the LVQuan19 dataset with reference segmentations for all time instants. Summary of Results: The method achieved detection performance with F1-score values up to 0.99 on real data, particularly for severe errors. Regarding the correction step, the selected strategy effectively improved segmentation consistency, achieving Dice coefficient values close to 0.95, indicating excellent agreement with reference segmentations. Contributions and Impact to IS area: This study contributes to the Information Systems field by introducing a method that improves the reliability of automated computer-aided systems for diagnosis. Although validated in the context of LV segmentation, the proposed approach can be applied to other domains where sequential data consistency is critical.

Ähnliche Arbeiten