Source-linked AI summary
End-to-end Prostate Cancer Detection in bpMRI via 3D CNNs: Effects of Attention Mechanisms, Clinical Priori and Decoupled False Positive Reduction
Anindo Saha, Matin Hosseinzadeh, Henkjan Huisman
TL;DR
The paper tackles automated csPCa localization in bpMRI despite lesion heterogeneity and benign mimics. It develops an end-to-end 3D CAD system combining attention-based detection, decoupled residual false-positive reduction, and a probabilistic anatomical prior. The system outperformed four recent 3D baselines on institutional testing and showed moderate agreement with expert radiologists and pathologists on external biopsy-confirmed scans.
Problem
csPCa lesions can resemble benign pathology, while biopsy and MRI interpretation have limitations that motivate reliable automated detection.
Method
The system combines an attention-based 3D detection network, a decoupled residual classifier for false-positive reduction, and an anatomical prior encoding csPCa spatial prevalence and zonal distinction.
Results
The CAD system significantly outperformed four recent 3D U-Net adaptations, while the classifier reduced false positives by up to 12.89% per patient using less than 1% of total system parameters.
Takeaways & Limitations
The study supports 3D CAD trained with radiologically estimated annotations as a potential second reader for csPCa detection across institutional and external testing datasets.
Takeaways & Limitations
The authors state that CNNs remain inadequate as stand-alone solutions, despite CAD showing moderate agreement with clinical experts.
Abstract
from arXiv · showhide
We present a multi-stage 3D computer-aided detection and diagnosis (CAD) model for automated localization of clinically significant prostate cancer (csPCa) in bi-parametric MR imaging (bpMRI). Deep attention mechanisms drive its detection network, targeting salient structures and highly discriminative feature dimensions across multiple resolutions. Its goal is to accurately identify csPCa lesions from indolent cancer and the wide range of benign pathology that can afflict the prostate gland. Simultaneously, a decoupled residual classifier is used to achieve consistent false positive reduction, without sacrificing high sensitivity or computational efficiency. In order to guide model generalization with domain-specific clinical knowledge, a probabilistic anatomical prior is used to encode the spatial prevalence and zonal distinction of csPCa. Using a large dataset of 1950 prostate bpMRI paired with radiologically-estimated annotations, we hypothesize that such CNN-based models can be trained to detect biopsy-confirmed malignancies in an independent cohort. For 486 institutional testing scans, the 3D CAD system achieves 83.69$\pm$5.22% and 93.19$\pm$2.96% detection sensitivity at 0.50 and 1.46 false positive(s) per patient, respectively, with 0.882$\pm$0.030 AUROC in patient-based diagnosis $-$significantly outperforming four state-of-the-art baseline architectures (U-SEResNet, UNet++, nnU-Net, Attention U-Net) from recent literature. For 296 external biopsy-confirmed testing scans, the ensembled CAD system shares moderate agreement with a consensus of expert radiologists (76.69%; $kappa$ $=$ 0.51$\pm$0.04) and independent pathologists (81.08%; $kappa$ $=$ 0.56$\pm$0.06); demonstrating strong generalization to histologically-confirmed csPCa diagnosis.
1. Introduction
The paper addresses the difficulty of detecting clinically significant prostate cancer (csPCa) amid heterogeneous lesions and visually similar benign tissue. It proposes a 3D CAD system combining attention, decoupled false-positive reduction, and an anatomical prior, evaluated across institutional and external datasets.
- csPCa ranges from indolent tumors to aggressive malignancies, making accurate diagnosis clinically important.
- MRI interpretation is challenged by multifocal, variable-shaped csPCa lesions resembling benign conditions and by low inter-reader agreement without experienced radiologists.
- Prior work suggests anatomical information can improve sensitivity, while most existing csPCa CAD systems operate on 2D images because of computational and imaging constraints.
- The proposed system harmonizes a 3D detection network with dual attention, a residual patch-wise classifier for decoupled false-positive reduction, and a probabilistic anatomical prior.
- The study evaluates whether radiologically estimated annotations can support detection of histologically confirmed malignancies using 486 institutional and 296 independent scans.
2.1. Dataset
The study uses retrospective prostate bpMRI cohorts with radiological annotations at RUMC and biopsy-supported external evaluation at ZGT. Data include zonal prostate segmentations used to construct and align an anatomical prior.
- 2436 MRI studies from 2317 RUMC patients were acquired between January 2016 and January 2018.
- 1584, 366, and 486 RUMC scans were assigned to training, validation, and testing sets using double-stratified patient-level sampling.
- 296 ZGT scans from 296 patients formed an additional external cohort acquired between March 2015 and January 2017.
- ZGT cases underwent TRUS-guided biopsy, with MRI-guided biopsy for suspicious lesions and independent uropathologist review.
- RUMC and ZGT contained 909 and 86 malignant cases, respectively, with 1092 and 97 csPCa lesions.
- Zonal segmentations achieved DSC values of 0.90±0.01 for whole gland, 0.85±0.02 for TZ, and 0.63±0.03 for PZ, supporting anatomical-prior construction.
2.2. Model Architecture
The CAD architecture combines two parallel 3D CNNs with decision fusion: an attention-based detector uses anatomical priors, while a classifier identifies false-positive regions to improve specificity.
- The system comprises parallel 3D CNNs M1 and M2 followed by a decision fusion node that produces one detection map.
- M1 generates preliminary voxel-level csPCa detections using a 144×144×18-voxel ROI that extends beyond the prostate.
- M1 combines channel-wise SE attention with grid-attention gates to emphasize discriminative feature dimensions and spatially important prostatic structures.
- M2 independently classifies overlapping prostate-centered patches and supplies malignancy scores for false-positive reduction.
- A malignant patch label requires at least a specified percentage of malignant voxels, addressing noise from voxel-to-patch label conversion.
- NDF aggregates M1 and M2 predictions to retain M1 sensitivity while reducing benign-region false positives.
- An aligned probabilistic anatomical prior is concatenated as an additional M1 input channel to encode spatial prevalence and zonal distinction of csPCa.
2.3. Experimental Design
Experiments evaluate patient-based diagnosis and volumetric detection under consistent preprocessing and training conditions. Performance uncertainty is estimated through cross-validation or bootstrap replication.
- Candidate systems used identical preprocessing, augmentation, tuning, and train-validation pipelines within each comparison experiment.
- Patient-based diagnosis performance was evaluated using receiver operating characteristic analysis.
- A 0.10 DSC threshold was selected because csPCa lesions are often small, have indistinct margins, and show substantial inter-reader variability.
- Metrics were computed in 3D across complete image volumes.
- Confidence intervals used twice the standard deviation from five-fold cross-validation for validation sets or 1000 bootstrap replications for testing sets.
3. Results and Analysis
Across classification and detection experiments, the proposed 3D CAD system improved false-positive control and maintained strong sensitivity on institutional and external testing sets. The anatomical prior supported generalization, while ensembling delivered only small gains at substantially higher computational cost.
- Architecture and label noise: SEResNet achieved the highest AUROC across training schemes, while patch-wise supervision outperformed whole-image training and increasing τ improved patch-level AUROC by up to 10%.At τ =0.1%, no patient-level label swaps occurred, unlike τ values of 0.5% and 1.0%.
- Ensembling and clinical comparison: CAD∗ improved patient-based AUROC by only 0.3–0.4% across testing datasets while requiring 2.6× trainable parameters, 2.5× VRAM and 1.3× inference time.On TS2, CAD∗ shared 76.69% agreement with expert radiologists and 81.08% agreement with independent pathologists.
4. Discussion and Conclusion
The study concludes that attention, decoupled false-positive reduction, and an anatomical prior jointly support 3D csPCa detection, localization, and cross-dataset generalization. The system shows promise as a second reader, but vendor-specific acquisition limits broader generalization.
- Attention detection network: M1’s dual-attention network detected more csPCa lesions than four adapted 3D U-Net architectures at 1.00 false positive per patient.It also achieved the highest detection sensitivity across testing datasets between 0.10–2.50 false positives per patient.
- Decoupled false-positive reduction: M2 reduced false positives by up to 12.89% while retaining maximum detection sensitivity and using less than 1% of the 3D CAD system’s parameters.Its independently trained, decoupled design also permits modular tuning, upgrading, or replacement.
- Clinical anatomical prior: The anatomical prior contributed most to 3D CAD generalization across multi-institutional testing datasets while adding negligible trainable parameters.Similar experiments with M2 produced no statistical improvement, consistent with its spatially invariant patch-wise classification role.
- Generalization: Training on 1950 scans with radiologically estimated annotations was intended to support detection of csPCa in an independent biopsy-confirmed cohort representing diverse clinical workflow studies.The external testing set included cases beyond the training annotations’ source distribution.
- Clinical relevance and limitations: CAD∗ showed moderate agreement with clinical experts and was proposed as a potential second reader, although the authors state that CNNs remain inadequate as stand-alone solutions.External examples also show errors on smaller lesions amid severe inflammation or fibrosis from benign conditions.
- Clinical relevance and limitations: Generalization remains constrained because all prostate scans were acquired using MRI scanners from the same vendor.The authors identify domain adaptation as a potential measure for heterogeneous acquisition conditions.