A multi-object deep neural network architecture to detect prostate anatomy in T2-weighted MRI: Performance evaluation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Prostate gland segmentation is crucial for estimating gland volume and managing prostate diseases.
• Accurate segmentation aids in diagnosis, treatment planning, and monitoring of prostate conditions.

Purpose of the Study:

• To develop and evaluate a 2D-3D convolutional neural network (CNN) ensemble, PPZ-SegNet, for automated prostate gland and peripheral zone segmentation.
• To assess the model's performance on diverse public datasets using T2-weighted MRI sequences.

Main Methods:

• A 2D-3D CNN ensemble (PPZ-SegNet) was developed using Bayesian hyperparameter optimization.
• The model was trained on 150 T2W MRI prostate cases and validated using five-fold cross-validation.
• Performance was evaluated on four independent test cohorts (283 cases total) using Dice similarity coefficient and Hausdorff distance.

Main Results:

• PPZ-SegNet achieved average Dice scores of 0.86 (Test #1), 0.79 (Test #2), 0.81 (Test #3), and 0.62 (Test #4).
• Segmentation performance showed improvement with larger prostate volumes in three of the four test cohorts.
• Variations in Dice scores across cohorts indicate a need for more generalized models.

Conclusions:

• The developed PPZ-SegNet demonstrates effective automated segmentation of the prostate and peripheral zone in T2W MRI.
• The study highlights the importance of model diversity to accommodate variations in gland size and other factors for universal segmentation.
• Further development is needed to create a robust, universal network for comprehensive prostate segmentation.