PhD Scientific Days 2023

Clinical Medicine - Posters I

Pre-embolization Volumetry of Uterine Leiomyomas using Self-adapting Machine Learning Frameworks

Előadó neve

Dr. Fazekas, Szuzina

Neptun code

JZINDU

Előadó munkahelye

Semmelweis University Medical Imaging Centre

Előadó telefonszáma

+36300981870

Előadó e-mail címe

szuzina1997@gmail.com

Az előadás címe

Pre-embolization Volumetry of Uterine Leiomyomas using Self-adapting Machine Learning Frameworks

Szerző(k) neve és munkahelye

Szuzina Fazekas 1, Robert Stollmayer 1, Bettina Katalin Budai 1, Kolos Turtoczki 1, Ildiko Kalina 1, Novak Pal Kaposi 1, Pal Maurovich-Horvat 1, Viktor Berczi 1
1 Semmelweis University, Medical Imaging Centre, Budapest

Bemutatás módja

Poszter

Szekció

Clinical Medicine - Posters I

Language of the presentation

English

Preferred session

Clinical Medicine

Összefoglaló szövege

Introduction: Uterine fibroid embolization (UFE) is a minimally invasive alternative to hysterectomy. Artificial intelligence (AI) algorithms, such as convolutional neural networks (CNN), have been successfully applied to the automated measurement of tumour volumes. In contrast, in the field of fibroid volumetry results are scarce, as AI algorithms require time and resources to be properly tuned.

Aims: We aimed to assess the potential of two self-adapting CNN frameworks (2D and 3D nnU-Net) for performing fully automated uterine fibroid segmentation.

Method: 130 axial pelvic T2-weighted SPAIR MRI scans of UFE patients were retrospectively retrieved, and the fibroids were manually segmented. The fibroid volumes and segmentation maps were processed with nnU-Net: first, the algorithm determined the optimal preprocessing, network architecture, training, and postprocessing steps, afterwards, it performed training via 5-fold cross-validation (CV), the training/validation ratio for 48 images was 4:1. After each round of CV, the trained models were combined into an ensemble, and evaluated on a hold-out test set of 40 cases. The overlap between the predicted and manually segmented volumes were measured by calculating the Dice scores (DSC), Jaccard-coefficients (JC), and average Hausdorff distances (aHD).

Results: The difference between the fibroid volume calculated from the manually segmented masks and the masks predicted by 2D nnU-Net were -14,23±56,13 cm3 (mean±SD), by 3D nnU-Net were 15,11±39,27 cm3; Pearson’s correlation coefficient was 0.975 and 0.990; DSC of 0,764±0,182 and 0,854±0,128; Jaccard-coefficient of 0,647±0,201 and 0,756±0,141; and aHD of 1,649±2,683 and 1,002±2,379 mm were achieved.

Conclusions: State-of-the-art machine learning methods, such as 3D nnU-Net, are well applicable for automated volume measurement of uterine fibroids despite the significant variations in size and number.

Funding: The results publicised in this work was reached with the sponsorship of Gedeon Richter Talentum Foundation in framework of Gedeon Richter Excellence PhD Scholarship of Gedeon Richter.

University and Doctoral School

Semmelweis University, Doctoral School of Theoretical and Translational Medicine

Supervisor

Prof. Dr. Viktor Bérczi, Dr. Novák Pál Kaposi

Publication of my abstract

I give consent to the publication of my abstract on the website of the congress.

Kind

Szabad

Status

elfogadva

Accepted presentation method

poszter

Előadás fájl jóváhagyás

nem rendelkezett róla

Előadó

7239

Start

11:48

End

11:53

Authors (legacy)

Szuzina Fazekas 1, Robert Stollmayer 1, Bettina Katalin Budai 1, Kolos Turtoczki 1, Ildiko Kalina 1, Novak Pal Kaposi 1, Pal Maurovich-Horvat 1, Viktor Berczi 1
1 Semmelweis University, Medical Imaging Centre, Budapest