2,281 posters, 63 videos, 56 audios, 43 topics, 18 sessions, 6,174 authors, 2,323 institutions
ePostersLive by SciGen Technologies S.A. All rights reserved.
RCOG World Congress 2025
June 23 - 25, 2025 | ExCeL London, UK
PP.0003
Evaluating the impact of measurement calipers on deep learning-based ovarian cancer diagnosis
Poster Presenter
Authors
Axel Geysels, Giulia Garofalo, Stefan Timmerman, Francesca Buonomo, Daniela Fischerová, Antonia Carla Testa, Francesca Moro, Lil Valentin, Povilas Sladkevicius, Ligita Jokubkiene, Marek Kudla, Caroline Van Holsbeke, Artur Czekierdowski, Elisabeth Epstein, Yvette Groszmann, Bart De Moor, Ben Van Calster, Wouter Froyman, Dirk Timmerman
Affiliations
Karolinska Institutet, KU Leuven, IRCCS Maternal and Infant Burlo Garofolo, Charles University and General University Hospital, Fondazione Policlinico Universitario Agostino Gemelli IRCCS, Skåne University Hospital, Medical University of Silesia, Ziekenhuis Oost-Limburg, Medical University of Lublin, Harvard Medical School
Part of Topic
AI & digital health
Evaluating the impact of measurement calipers on deep learning-based ovarian cancer diagnosis: a prospective, multicentre analysis
Axel Geysels, Giulia Garofalo, Stefan Timmerman, Francesca Buonomo, Daniela Fischerová, Antonia Carla Testa, Francesca Moro, Lil Valentin, Povilas Sladkevicius, Ligita Jokubkiene, Marek Kudla, Caroline Van Holsbeke, Artur Czekierdowski, Elisabeth Epstein, Yvette Groszmann, Bart De Moor, Ben Van Calster, Wouter Froyman, Dirk Timmerman
Objective: To determine whether measurement calipers in transvaginal ultrasound images can bias the performance of AI models in ovarian-cancer diagnosis
Design:
Creation of two datasets:
-Raw (without calipers)
-Overlay (with calipers)
Assess four train → test scenarios (Raw→Raw, Overlay→Overlay, Raw→Overlay, Overlay→Raw):
Upper bound AUC (95% CI)
Evaluated on
Raw
Overlay
Trained on
Raw
0.909 (0.893-0.925)*
0.907 (0.891-0.923)
Overlay
0.913 (0.899-0.928)
0.914 (0.900-0.929)*
*p=0.26 (paired DeLong)
Results (qualitative):
Discussion:
-AUC: no significant change in training with or without calipers.
-Saliency: heatmap shifts reflect training noise more than caliper cues.
-Future research: validate using images from machines of different ultrasound vendors.