Title:A Light, 3D UNet-based Architecture for fully Automatic Segmentation of
Prostate Lesions from T2-MRI Images
Volume: 20
Author(s): Larisa-Gabriela Coroama, Laura Diosan, Teodora Telecan*, Iulia Andras, Nicolae Crisan, Anca Andreica, Cosmin Caraiani, Andrei Lebovici, Zoltán Bálint*Bianca Boca
Affiliation:
- Department of Urology, Faculty of Medicine, Iuliu Hațieganu University of Medicine and Pharmacy, 400012 Cluj-Napoca, Romania
- Department of Urology, Municipal Clinical Hospital, 400139 Cluj-Napoca, Romania
- Department of Biomolecular Physics, Faculty of Physics, Babeș-Bolyai University, 400084 Cluj-Napoca, Romania
Keywords:
Fully automatic 3D image segmentation, Prostate cancer characterization., T2 MRI images, Slim 3D UNet, Convolutional neuronal network, UNet architecture.
Abstract:
Introduction:
Prostate magnetic resonance imaging (MRI) has been recently integrated into the pathway of diagnosis of prostate cancer (PCa). However, the lack
of an optimal contrast-to-noise ratio hinders automatic recognition of suspicious lesions, thus developing a solution for proper delimitation of the
tumour and separating it from the healthy parenchyma are extremely important.
Methods:
As a solution to this unmet medical need, we aimed to develop a decision support system based on artificial intelligence, which automatically
segments the prostate and any suspect area from the 3D MRI images.
We assessed retrospective data from all patients diagnosed with PCa by MRI-US fusion prostate biopsy, who underwent prostate MRI in our
department due to a clinical or biochemical suspicion of PCa (n=33). All examinations were performed using a 1.5 Tesla MRI scanner. All images
were reviewed by two radiologists, who performed manual segmentation of the prostate and all lesions. A total of 145 augmented datasets were
generated. The performance of our fully automated end-to-end segmentation model based on a 3D UNet architecture and trained in two learning
scenarios (on 14 or 28 patient datasets) was evaluated by two loss functions.
Results:
Our model had an accuracy of over 90% for automatic segmentation of prostate and PCa nodules, as compared to manual segmentation. We have
shown low complexity networks, UNet architecture with less than five layers, as feasible and to show good performance for automatic 3D MRI
image segmentation. A larger training dataset could further improve the results.
Conclusion:
Therefore, herein, we propose a less complex network, a slim 3D UNet with superior performance, being faster than the original five-layer UNet
architecture.
