Segmentation of prostate and prostate zones using deep learning : A multi-MRI vendor analysis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate segmentation of the prostate and its peripheral zone (PZ) is crucial for diagnosis and treatment planning.
• Current MRI segmentation methods may lack generalizability across different imaging vendors.

Purpose of the Study:

• To develop a deep-learning algorithm for reliable prostate and PZ segmentation across multiple MRI vendors.
• To evaluate the performance of a multistream 3D convolutional neural network for this task.

Main Methods:

• A retrospective study utilized 550 T2-weighted MRIs from Siemens and GE vendors.
• A multistream 3D convolutional neural network was trained for automatic segmentation of the prostate and PZ.
• Models were trained on vendor-specific and combined datasets, with Dice coefficient (DSC) used for evaluation.

Main Results:

• The combined model achieved robust prostate segmentation DSCs of 0.893 ± 0.036 (Siemens) and 0.825 ± 0.112 (GE).
• For PZ segmentation, the combined model yielded DSCs of 0.811 ± 0.079 (Siemens) and 0.788 ± 0.093 (GE).
• Vendor-specific models showed decreased performance when tested on data from the other vendor, highlighting the benefit of combined training.

Conclusions:

• The developed deep-learning network demonstrates performance comparable to interexpert variability for prostate and PZ segmentation.
• Training the network on combined data from multiple MRI vendors is essential for creating a universal segmentation model.
• This approach enhances the reliability and applicability of automated prostate segmentation in clinical settings.