Multiparametric Ultrasound Imaging of Prostate Cancer Using Deep Neural Networks

Area of Science:

• Medical imaging
• Artificial intelligence in medicine
• Prostate cancer diagnostics

Background:

• Accurate prostate cancer detection is crucial for effective treatment.
• Multiparametric ultrasound (mpUS) combines multiple imaging modalities for enhanced visualization.
• Deep neural networks (DNNs) show promise in improving medical image analysis.

Purpose of the Study:

• To train a DNN to generate mpUS volumes for prostate cancer detection.
• To evaluate the performance of the DNN-generated mpUS volumes compared to existing methods.
• To identify key ultrasound modalities contributing to the DNN model.

Main Methods:

• A DNN was trained using co-registered acoustic radiation force impulse (ARFI) imaging, shear wave elasticity imaging (SWEI), quantitative ultrasound-midband fit (QUS-MF), and B-mode data.
• Data were acquired from 15 men with biopsy-confirmed prostate cancer prior to radical prostatectomy.
• Histopathology data was used for manual segmentation of index lesions and non-cancerous regions.

Main Results:

• The DNN significantly increased lesion contrast-to-noise ratio (CNR) compared to a linear support vector machine (SVM) in both phantom and in vivo studies (2.79 ± 0.88 vs. 1.98 ± 0.73, p < 0.001).
• DNN-generated mpUS volumes clearly depicted histopathology-confirmed prostate cancer.
• Omitting input modalities reduced CNR, highlighting the importance of stiffness and echogenicity-based data.

Conclusions:

• A DNN can be optimized to generate high-CNR mpUS prostate volumes from multiple ultrasound modalities.
• The DNN approach demonstrates superior performance in prostate cancer detection compared to a linear SVM.
• This AI-driven method enhances the diagnostic capability of mpUS for prostate cancer.