Original scientific paper

https://doi.org/10.7906/indecs.24.6.9

The Role of Multidisciplinary Expertise in AI Training Data: Quantifying Variability in Prostate and Urethral MRI Segmentation

Szilvia Totin ; University of Szeged – Department of Radiology, Szeged, Hungary
Andras Negyessy ; University of Szeged – Department of Urology, Szeged, Hungary
William Holmlund ; Umeå University – Department of Diagnostics and Intervention, Umeå, Sweden
Attila Simko ; Umeå University – Department of Diagnostics and Intervention, Umeå, Sweden
Istvan Elod Kiraly ; University of Szeged – Department of Urology, Szeged, Hungary
Adam Hodoniczki ; University of Szeged – Department of Urology, Szeged, Hungary
Benedek Danka ; University of Szeged – Department of Radiology, Szeged, Hungary
Eniko Koos ; University of Szeged – Department of Radiology, Szeged, Hungary
Viktor Rogowski ; Skåne University Hospital – Department of Hematology, Oncology, and Radiation Physics, Lund, Sweden
Christian Jamtheim Gustafsson ; Lund University – Department of Translational Medicine, Medical Radiation Physics, Malmö, Sweden
Miklos Zrinyi ; University of Szeged Faculty of Health Sciences and Social Studies, Department of Nursing, Szeged, Hungary
Angelika Szatmari ; University of Szeged Faculty of Health Sciences and Social Studies, Department of Nursing, Szeged, Hungary
Peter Palasti ; University of Szeged – Department of Radiology, Szeged, Hungary
Linda Varga ; University of Szeged – Department of Oncotherapy, Szeged, Hungary
Tufve Nyholm ; Umeå University – Department of Diagnostics and Intervention, Umeå, Sweden
Zoltan Bayori ; University of Szeged – Department of Urology, Szeged, Hungary
Zsuzsanna Fejes ; University of Szeged – Department of Radiology, Szeged, Hungary *

* Corresponding author.

Abstract

Introduction: Accurate segmentation of prostate mpMRI scans is essential for oncological therapy planning and the development of AI-based decision support systems. The aim of our study was to validate the annotation reliability of observer groups with different levels of professional experience in prostate and urethral segmentation to establish an expert AI database.
Methods: A total of 35 mpMRI cases were processed using 3D Slicer software. Seven participants were categorized in three groups: uroradiologists, radiology and urology residents, and medical students. Interobserver variability was quantified using Dice Similarity Coefficient (DSC), Surface Dice (at 1 mm and 3 mm tolerances), and Central Line Distance (CLD) metrics, with validation performed via Dunn’s post-hoc test.
Results: Prostate DSC values were statistically equivalent between uroradiologists, urologists, and residents (p_holm = 1,000; z = 0,543). Midline determination of the urethra (CLD) showed
specialty-specific differences between radiologists and urologists (p = 0,039), with an average difference of 3,3-3,7 mm (with min 2,9 mm; max 4,4 mm). While the 1 mm surface precision of medical students was significantly lower than that of experts (z = 4,639; p_holm < 0,001), no significant difference was observed within the 3 mm clinical tolerance margin. Notably, certain student subgroups reached expert-level performance even in urethra segmentation (p_holm = 1,000).
Conclusion: Residents and uroradiologists from related fields generate expert-level annotation. The involvement of trained non-experts (students) represents a valid crowdsourcing alternative, provided that quality assurance is aligned with the 3 mm clinical tolerance threshold.

Keywords

segmentation; interobserver variability; artificial intelligence; prostate mpMRI; 3D slicer

Hrčak ID:

350262

URI

https://hrcak.srce.hr/350262

Publication date:

30.12.2026.

Visits: 0 *