Prostate Cancer Detection
Using Zonal-Aware Attention-Guided U-Net
We developed an advanced deep learning solution for prostate cancer detection using bi-parametric MRI (bpMRI) with zonal awareness integrated into an Attention-Guided U-Net. This approach improves diagnostic accuracy by incorporating prostate anatomy knowledge, enabling precise segmentation and classification of clinically significant lesions.
What We Built
A comprehensive AI diagnostic pipeline, including:
- Attention-Guided U-Net architecture for prostate zone and lesion segmentation
- Zonal awareness integration leveraging anatomical knowledge of prostate regions
- bpMRI data processing for multi-parametric medical imaging
- Dual-model system for segmentation and lesion detection
- Five-fold cross-validation for robust model evaluation
The system acts as an intelligent diagnostic assistant, dynamically segmenting prostate zones and detecting significant lesions with higher accuracy than baseline models.
Key Features
Zonal-Aware Segmentation
- Integration of anatomical knowledge for prostate zone identification
- Attention gates focus on critical regions during decoding
- Enhanced detection of clinically significant lesions
Dual-Model Architecture
- Separate segmentation and lesion detection models
- Shared U-Net backbone with attention mechanisms
- Precise localization of cancerous areas
Comprehensive Training Pipeline
- 5:1 training-testing split for segmentation
- Five-fold cross-validation for lesion detection
- Dice coefficient loss optimization with Adam optimizer and adaptive learning rate
Multi-Model Benchmarking
- Comparison with pre-trained 3D CNNs: ResNet201, VGG19, SEResNet152
- Performance evaluated using Average Precision, AUROC, Dice, IoU, and Jaccard metrics
How It Works
bpMRI scans are ingested and preprocessed for model input
Prostate segmentation model identifies zones using attention-guided U-Net
Zonal awareness guides lesion detection to relevant regions
Lesion detection model classifies suspicious areas
Attention gates prioritize key features during decoding
Results evaluated using Dice similarity, IoU, and Jaccard metrics
Results
Training Strategy
Epoch 15 for segmentation, up to 200 epochs for lesion detection
Segmentation Accuracy
Dice 0.70–0.75, IoU 0.55–0.61, Jaccard 0.55–0.60 at optimal threshold 0.7
Performance Comparison
Outperforms 3D ResNet201, VGG19, and SEResNet152 baselines
Best Model
Attention-U-Net with batch normalization and dropout
Value Dilevered
This project advanced prostate cancer detection by combining deep learning with clinical anatomical knowledge. The zonal-aware Attention-Guided U-Net enables earlier, more accurate detection of significant lesions while reducing false positives, demonstrating the power of integrating domain expertise with AI for patient-specific diagnostics.
Want to build AI-powered Problem solutions?
#Deep Learning #Medical Imaging #Prostate Cancer Detection #Attention-Guided U-Net #bpMRI #Computer-Aided Diagnosis #Healthcare AI