Algorithms for classification of sequences and segmentation of prostate gland: an external validation study

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Oncology

Background:

• Prostate cancer diagnosis relies heavily on magnetic resonance imaging (MRI).
• Accurate segmentation and classification of prostate MRI are crucial for effective diagnosis and treatment planning.
• Current manual methods can be time-consuming and subject to inter-observer variability.

Purpose of the Study:

• To externally validate two artificial intelligence (AI) models for prostate multiparametric MRI (mpMRI) sequence classification and prostate gland segmentation on T2-weighted imaging (T2WI).
• To assess the performance of AI models across different MR field strengths and sequences.

Main Methods:

• Retrospective collection of mpMRI data from 719 patients across two hospitals, using nine MR scanners from four vendors.
• Utilized a pretrained Med3D deep learning architecture for image classification and a UNet-3D architecture for prostate segmentation.
• Evaluated segmentation accuracy using Dice Similarity Coefficient (DSC), Volume Similarity (VS), and Average Hausdorff Distance (AHD).

Main Results:

• The classification model achieved 99% accuracy with a kappa of 0.932.
• Segmentation model performance showed median DSC of 0.942-0.955, median VS of 0.974-0.982, and median AHD of 5.55-6.49 mm.
• Model performance varied significantly across different magnetic field strengths (1.5 T, 3.0 T), with higher accuracy at 3.0 T.

Conclusions:

• AI models for mpMRI classification and prostate segmentation demonstrate strong performance in external validation.
• These validated AI tools have the potential to enhance efficiency in prostate volume measurement and cancer detection using mpMRI.
• The models can significantly improve workflow efficiency for cancer detection, prostate volume measurement, and image-guided biopsies.