Calculation of Apparent Diffusion Coefficients in Prostate Cancer Using Deep Learning Algorithms: A Pilot Study.

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Apparent diffusion coefficients (ADCs) from diffusion-weighted imaging (DWI) are crucial for prostate cancer management.
• DWI quality is often compromised by distortions and artifacts, limiting ADC accuracy and reproducibility.
• Current methods to improve DWI quality are resource-intensive, hindering clinical application.

Purpose of the Study:

• To develop and assess a supervised learning framework using generative adversarial networks (GANs) for synthesizing ADC images.
• To create a method for generating accurate ADC maps without reliance on specific MRI hardware or extended scan times.

Main Methods:

• A prospective study involved 200 patients with suspected prostate cancer and 10 healthy volunteers.
• Full field-of-view (FOV) DWI (f-DWI) and zoomed-FOV DWI (z-DWI) were acquired with b-values of 50, 1000, and 1500 s/mm².
• A GAN-based method synthesized ADC (s-ADC) values from f-DWI (b=1000) using z-ADC as a reference, evaluating image quality, distortion, reproducibility, and diagnostic performance.

Main Results:

• Synthesized ADC (s-ADC) maps at b=1000 demonstrated superior image quality (higher PSNR, SSIM, FSIM) and lower error (RMSE) compared to other synthesized b-values.
• Both z-ADC and s-ADC (b=1000) exhibited reduced distortion and enhanced quantitative reproducibility.
• s-ADC (b=1000) and z-ADC showed improved tumor detection and classification performance over f-ADC.

Conclusions:

• Deep learning-based ADC map generation is a viable alternative to traditional z-ADC maps.
• This GAN-driven approach offers a hardware-independent and time-efficient solution for improved prostate cancer imaging.