Source-linked AI summary
Prostate Cancer Detection using Deep Convolutional Neural Networks
Sunghwan Yoo, Isha Gujrathi, Masoom A. Haider, Farzad Khalvati
TL;DR
Clinically significant prostate cancer detection remains challenging because current diagnostic pathways can involve PSA-related over-diagnosis and radiologist performance on DWI varies. This paper develops an automated CNN-based pipeline for slice- and patient-level classification, achieving AUCs of 0.87 and 0.84, respectively.
Problem
Current prostate cancer diagnosis can involve PSA-related over-diagnosis, while radiologists detecting PCa on DWI have reported ROC AUCs ranging from 0.69 to 0.81.
Method
The paper proposes an automated two-level PCa classification pipeline combining five individually trained CNNs, a decision tree-based feature extractor, and a Random Forest classifier, evaluated with separate training, validation, and test sets.
Results
The pipeline achieved a DWI slice-level AUC of 0.87 (95% CI: 0.84−0.90) and a patient-level AUC of 0.84 (95% CI: 0.76−0.91).
Takeaways & Limitations
The proposed slice-based approach avoids ROI generation and uses CNN-derived features for patient-level PCa classification.
Takeaways & Limitations
ROI-based limitations motivate the approach, but the pipeline still requires user indication of the first and last prostate-containing slices.
Abstract
from arXiv · showhide
Prostate cancer is one of the most common forms of cancer and the third leading cause of cancer death in North America. As an integrated part of computer-aided detection (CAD) tools, diffusion-weighted magnetic resonance imaging (DWI) has been intensively studied for accurate detection of prostate cancer. With deep convolutional neural networks (CNNs) significant success in computer vision tasks such as object detection and segmentation, different CNNs architectures are increasingly investigated in medical imaging research community as promising solutions for designing more accurate CAD tools for cancer detection. In this work, we developed and implemented an automated CNNs-based pipeline for detection of clinically significant prostate cancer (PCa) for a given axial DWI image and for each patient. DWI images of 427 patients were used as the dataset, which contained 175 patients with PCa and 252 healthy patients. To measure the performance of the proposed pipeline, a test set of 108 (out of 427) patients were set aside and not used in the training phase. The proposed pipeline achieved area under the receiver operating characteristic curve (AUC) of 0.87 (95% Confidence Interval (CI): 0.84-0.90) and 0.84 (95% CI: 0.76-0.91) at slice level and patient level, respectively.
Introduction
The paper motivates automated prostate cancer detection from DWI because current diagnosis has over-diagnosis and inter-observer challenges, then proposes a two-level CNN-based classification pipeline.
- Motivation: Early prostate cancer detection matters because the disease progresses slowly and early-stage survival rates are high.
- Clinical challenges: PSA screening can lead to over-diagnosis, unnecessary needle biopsies, and potential over-treatment.
- Technical background: Deep CNNs extract hierarchical image features and have shown promising results in classification, segmentation, and object detection.
- Contribution: The proposed work performs automated PCa classification at both slice and patient levels.
- Contribution: The pipeline combines five individually trained ResNet-inspired CNNs, decision-tree feature extraction, and a Random Forest classifier.
Results
The pipeline was evaluated with ROC and AUC on an independent test set at slice and patient levels. Its reported AUC was 0.87 for slice-level detection and 0.84 for patient-level classification.
- Evaluation: 108 patients and 1,486 slices formed the test set used for slice-level and patient-level evaluation.
- Slice-Level Performance: 0.87 slice-level AUC was achieved by the best CNN, with a 95% CI of 0.84−0.90.
- Slice-Level Performance: Table 1 compares slice-level performance across five individually trained deep CNNs.
- Patient-Level Performance: 0.84 patient-level AUC was achieved by the Random Forest classifier using features extracted through CNNs, with a 95% CI of 0.76−0.91.
Discussion
The discussion contrasts ROI-based and slice-based PCa detection, emphasizing automated cropping, independent evaluation, and CNN-derived patient-level classification.
- PCa MRI classification methods include radiomics-driven feature-based approaches and deep learning-based approaches.
- ROI-based CAD requires manual or automated segmentation, and segmentation errors can impair downstream PCa detection.
- ROI-based methods struggle to aggregate numerous regional predictions into acceptable patient-level classifications, often relying on simple voting.
- The proposed pipeline uses automatically center-cropped DWI images, requiring only the first and last prostate-containing slices instead of ROI generation.
- AUC 0.87 was achieved for slice-level detection using a test set of 108 patients or 1,486 slices.
- CNN outputs were used as features for patient-level PCa classification, while test data were isolated from training and validation for evaluation.
Data
The study retrospectively analyzed 427 biopsy-evaluated patients with PI-RADS scores of 3 or higher, defining clinically significant PCa using Gleason grade groups.
- 427 consecutive patients with PI-RADS scores of 3 or higher who underwent biopsy were included retrospectively.
- 175 patients had clinically significant PCa, while 252 patients were classified without clinically significant PCa.
- Clinically significant PCa was defined as Gleason score ≥7 or ISUP grade group GG≥2; GG=1 or GG=0 indicated no clinically significant PCa.
MRI Acquisition
DWI data were acquired on a 3T MRI system using multiple b values and transformed into six-channel CNN inputs after resizing and prostate-centered cropping.
- DWI was acquired between January 2014 and July 2017 using a Philips Achieva 3T whole-body MRI scanner.
- The transverse DWI sequence used b values of 0, 100, 400, and 1000 s mm−2, with 3 mm slices and a 140 × 140 matrix.
- DWI measures tissue sensitivity to Brownian motion and supports estimation of ADC and computation of high b-value images.
- DWI slices were resized to 144 × 144 pixels, center-cropped to 66 × 66 pixels, and represented using six channels.
Training, Validation, and Test sets
The 427-patient dataset was divided into training, validation, and test sets, with the test set reserved for evaluation and class ratios kept approximately balanced.
- 271 patients and 3,692 slices formed the training set.
- 48 patients and 654 slices formed the validation set, while 108 patients and 1,486 slices formed the test set.
- The training, validation, and test sets represented 64%, 11%, and 25% of the patients, respectively.
- The proportions of PCa and non-PCa patients were kept roughly similar across the datasets.
- Table 2 reports patient and slice counts with and without PCa for the training, validation, and test sets.
Data Preprocessing
The DWI images were normalized using a dataset-wide normalization function based on pixel values, the dataset mean, and standard deviation.
- Data Preprocessing: All DWI images were normalized across the entire dataset.The normalization used pixel values from individual MRI slices together with dataset-level statistics.
Pipeline
The proposed pipeline combines five CNNs, statistical feature extraction, feature selection, and Random Forest classification to produce slice-level and patient-level PCa results.
- Pipeline: The pipeline classifies each DWI slice with five individually trained CNN models.The CNN outputs feed later patient-level processing.
- Pipeline: First-order statistics are extracted from CNN probability outputs and selected using a decision tree-based feature selector.The extracted statistics include measures such as mean, standard deviation, and median.
- Pipeline: A Random Forest classifier uses the selected statistical features to classify patients with and without PCa.The classifier is trained and fine-tuned using validation-set features with 10-fold cross-validation.
- Pipeline: Each CNN receives a 66 × 66 × 6 MRI slice containing ADC, b0, b100, b400, b1000, and b1600 inputs.The pipeline outputs slice-level and patient-level results.
- Pipeline: The CNN architecture is based on ResNet, using convolutional layers and identity shortcut connections within residual blocks.When dimensions differ, a linear projection changes the shortcut dimension to match the output.
- Pipeline: A fully pre-activated residual network places batch normalization and ReLU before convolutional layers.The architecture also uses a three-layer bottleneck block to reduce training time without sacrificing performance.
A ResNet
The CNN system uses stochastic optimization and stacked generalization, aggregating probability-derived features before Random Forest patient classification.
- A ResNet: The CNNs were trained with stochastic gradient descent, batch size 8, dropout rate 0.90, weight decay 0.000001, and momentum 0.90.Binary cross entropy was used because the dataset was extremely unbalanced.
- Stacked Generalization: Randomness in CNN training can produce models that capture different features despite identical hyperparameters and input data.The pipeline exploits these differences through stacked generalization.
- Stacked Generalization: Performance significantly improved with five CNNs compared with a single CNN at the patient level.The five-CNN model achieved AUC 0.84 (CI: 0.76−0.91) versus AUC 0.71 (CI: 0.61−0.81), with 2-tailed P = 0.048.
- First Order Statistical Feature Extraction: Each patient generated 90 statistical features from CNN probability outputs, after which 26 best features were selected.The decision tree-based selector was trained and fine-tuned using 10-fold cross-validation on the validation set.
- Stacked Generalization: The Random Forest was trained on validation-set features and tested on the test set for patient-level classification.The reported testing procedure separates feature-model tuning from final evaluation.
- A ResNet: Training all five CNNs took 6 hours, while Random Forest training took less than 10 seconds and testing 108 patients took less than 1 minute.
Ethics approval and consent to participate
The retrospective single-institution studies received ethics approval, and the requirement for informed consent was waived.
- Ethics approval and consent to participate: The Sunnybrook Health Sciences Centre Research Ethics Boards approved the retrospective single-institution studies.The boards also waived the requirement for informed consent.