Textured-Based Deep Learning in Prostate Cancer Classification with 3T Multiparametric MRI: Comparison with...

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Standardized interpretation of prostate MRI requires significant expertise, leading to variability.
• Automated prostate cancer (PCa) classification can enhance MRI's diagnostic capabilities.
• Current methods like PI-RADS-based classification (PI-RADS-CLA) have limitations in accuracy and consistency.

Purpose of the Study:

• To evaluate a texture-based deep learning model (Textured-DL) for differentiating clinically significant PCa (csPCa) from non-csPCa.
• To compare the performance of Textured-DL against PI-RADS-CLA.

Main Methods:

• A cohort of 402 patients with 3T multiparametric MRI data was used, with data split for training, validation, and testing.
• Volumetric patches of T2-Weighted MRI and apparent diffusion coefficient images were input into a 3D CNN model (Textured-DL).
• Performance was compared against expert PI-RADS-CLA using sensitivity, specificity, and AUC, with statistical analysis including Mcnemar's test and bootstrapping.

Main Results:

• Textured-DL achieved a significantly higher AUC (0.85) compared to PI-RADS-CLA (0.73) for PCa classification (p < 0.05).
• Textured-DL demonstrated significantly higher specificity (0.70) than PI-RADS-CLA (0.47) (p < 0.05).
• Sub-analyses showed superior specificity for Textured-DL in the peripheral zone and for solitary tumors.

Conclusions:

• The texture-based deep learning model (Textured-DL) outperforms PI-RADS-CLA in classifying prostate cancer.
• Textured-DL shows improved specificity, particularly for peripheral zone lesions and solitary tumors.
• This AI model offers a promising advancement for more accurate and consistent prostate cancer assessment via MRI.