Deep Learning Regression for Prostate Cancer Detection and Grading in Bi-Parametric MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Prostate cancer (PCa) is a common malignancy in men.
• Bi-parametric MRI aids PCa diagnosis, but previous AI models focused on detection or classification separately.
• A simultaneous detection and grading approach is more clinically relevant.

Purpose of the Study:

• To develop and evaluate an end-to-end neural network for simultaneous PCa detection and Gleason Grade Group (GGG) grading from MRI.
• To investigate methods for encoding ordinal GGG information and incorporating prostate zone priors.
• To compare the proposed model's performance against standard classification and regression techniques.

Main Methods:

• A 2D U-Net architecture was employed, taking MRI slices as input and outputting lesion segmentation maps encoding GGG.
• A novel method for encoding ordinal GGG in the model target was proposed.
• Prostate zone segmentations and ensembling techniques were evaluated for performance enhancement.

Main Results:

• The model achieved a voxel-wise weighted kappa of 0.446 ±0.082 and a Dice score of 0.370 ±0.046 for clinically significant cancer segmentation.
• Lesion-wise weighted kappa on the test set was 0.13 ±0.27.
• The proposed GGG encoding method outperformed standard multiclass classification and multi-label ordinal regression.

Conclusions:

• The developed neural network effectively performs simultaneous detection and grading of prostate cancer from MRI.
• The novel GGG encoding strategy improves model performance for cancer aggressiveness assessment.
• This end-to-end approach offers a more clinically relevant solution for PCa diagnosis and grading.