[Paper Review] Encoding Clinical Priori in 3D Convolutional Neural Networks for Prostate Cancer Detection in bpMRI

Anindo Saha, Matin Hosseinzadeh+1 more|arXiv (Cornell University)|Oct 31, 2020

Prostate Cancer Diagnosis and Treatment24 references6 citations

TL;DR

This paper proposes encoding clinical priors into 3D U-Net architectures for improved prostate cancer detection in bi-parametric MRI (bpMRI). By constructing a probabilistic population prior based on zonal anatomy and lesion prevalence from 700 training scans, the method boosts patient-based diagnosis (up to 8.70% AUROC gain) and lesion-level detection (average 1.08 pAUC increase) across four 3D CNN architectures, even with limited, radiologist-estimated annotations.

ABSTRACT

We hypothesize that anatomical priors can be viable mediums to infuse domain-specific clinical knowledge into state-of-the-art convolutional neural networks (CNN) based on the U-Net architecture. We introduce a probabilistic population prior which captures the spatial prevalence and zonal distinction of clinically significant prostate cancer (csPCa), in order to improve its computer-aided detection (CAD) in bi-parametric MR imaging (bpMRI). To evaluate performance, we train 3D adaptations of the U-Net, U-SEResNet, UNet++ and Attention U-Net using 800 institutional training-validation scans, paired with radiologically-estimated annotations and our computed prior. For 200 independent testing bpMRI scans with histologically-confirmed delineations of csPCa, our proposed method of encoding clinical priori demonstrates a strong ability to improve patient-based diagnosis (upto 8.70% increase in AUROC) and lesion-level detection (average increase of 1.08 pAUC between 0.1-10 false positives per patient) across all four architectures.

Motivation & Objective

• To improve computer-aided detection of clinically significant prostate cancer (csPCa) in bpMRI using domain-specific clinical knowledge.
• To address limited and noisy training data by incorporating anatomical and prevalence priors into deep learning models.
• To evaluate whether explicit anatomical priors enhance performance across diverse 3D U-Net architectures in csPCa detection.
• To demonstrate that prior knowledge can improve generalization beyond radiologist-estimated annotations in histologically confirmed cases.

Proposed method

• A probabilistic population prior $P$ is constructed from 700 training bpMRI scans using radiologically-estimated csPCa annotations and CNN-generated prostate zonal segmentations.
• The prior encodes spatial likelihoods: 0.00 for background, $\mu$ for transitional zone (TZ), $3\mu$ for peripheral zone (PZ), and 1.00 for malignant voxels, with $\mu \in [0, 0.33]$ as a learnable hyperparameter.
• For each scan, the prior $P$ is aligned using case-specific prostate segmentations to ensure spatial correspondence without non-rigid registration.
• The prior is integrated into 3D U-Net, U-SEResNet, UNet++, and Attention U-Net architectures during training as an auxiliary supervision signal.
• Performance is evaluated on 200 independent testing scans with histologically confirmed csPCa delineations using AUROC and pAUC metrics.
• Ablation studies compare baseline models (no prior) with variants: prevalence prior ($\mu=0.00$), zonal prior ($\mu=0.33$), and hybrid prior ($\mu=0.01$).

Experimental results

Research questions

• RQ1 Can a data-driven, probabilistic population prior improve 3D CNN performance in csPCa detection on bpMRI?
• RQ2 How does incorporating zonal anatomy and lesion prevalence affect lesion-level and patient-based detection performance?
• RQ3 Does the prior enhance model robustness when training data is limited and annotations are radiologist-estimated?
• RQ4 How does the prior perform across diverse 3D U-Net architectures with varying attention and skip-connection mechanisms?

Key findings

• The proposed method improved patient-based diagnosis by up to 8.70% in AUROC across all four 3D CNN architectures when using the hybrid prior ($\mu=0.01$).
• Lesion-level detection showed an average pAUC increase of 1.08 between 0.1–10 false positives per patient across all FROC setups.
• The Attention U-Net achieved the highest AUROC of $0.790 \pm 0.066$ when using the hybrid prior, outperforming its baseline of $0.711 \pm 0.078$.
• The zonal prior ($\mu=0.33$) consistently outperformed the prevalence prior ($\mu=0.00$) in both AUROC and pAUC, indicating the value of anatomical zoning.
• All architectures benefited from prior integration, demonstrating robustness across different network designs and attention mechanisms.
• The method maintained performance gains even with limited training data and radiologist-estimated annotations, suggesting improved generalization to histologically confirmed cases.