TL;DR
The proposed HBR-Net-18 is a novel two-stage AI framework that significantly improves automated prostate cancer detection from quantitative Hybrid Multi-dimensional MRI (HM-MRI). By decoupling low-frequency intensity normalization (via a probabilistic Hadamard U-Net) from patch-level classification (via ResNet-18), the model prevents the classifier from exploiting spurious imaging artifacts, achieving a well-balanced patch-level accuracy of 93.8%.
What Problem Does This Paper Address?
While Magnetic Resonance Imaging (MRI) is essential for diagnosing prostate cancer (PCa), standard Convolutional Neural Networks (CNNs) often struggle with bias field inhomogeneities—low-frequency intensity variations caused by hardware limitations or patient anatomy.
When a deep learning model is trained end-to-end on raw MRI data, it frequently learns to exploit these spurious intensity variations rather than actual pathological features. This leads to poor generalization across different scanners and inconsistent clinical predictions. The authors address this challenge by treating prostate cancer detection as a bias-aware, decoupled two-stage learning problem.
Methods
The authors introduce HBR-Net-18 (Hadamard-Bias Network plus ResNet18), a modular framework operating on six tissue-based biomarkers derived from HM-MRI via a Physics-Informed Autoencoder (PIA).
The pipeline consists of two distinct stages:
-
Stage 1: Bias-Aware Biomarker Correction
A learnable probabilistic Hadamard U-Net is trained to suppress noise and correct low-frequency intensity inhomogeneities across the six biomarker maps. The Walsh–Hadamard Transform is highly efficient here because its energy compaction properties concentrate low-frequency bias fields into a small subset of coefficients. Crucially, once this stage converges, its parameters are frozen, strictly separating bias normalization from the downstream classification task. -
Stage 2: Patch-Level Cancer Classification
The corrected biomarker maps are focused on the prostate region and divided into overlapping $11 \times 11$ patches. To capture 3D spatial context, each target slice is stacked with its adjacent slices ($N-1, N, N+1$). These patches are upsampled through a learnable convolutional projection layer to match the input dimensions required by a ResNet-18 classifier, which outputs the final cancer probability.
Datasets
The framework was trained and evaluated using a dataset of 86 patients who underwent MRI on a 3T Philips scanner:
- 34 patients were cancer-free.
- 52 patients had confirmed prostate cancer.
- Each patient scan contained between 18 and 30 slices.
Ground-truth cancer and prostate masks were provided by expert radiologists. The model's performance was rigorously validated using a 5-fold cross-validation strategy split on a patient-by-patient basis to prevent data leakage.
Key Insights and Ideas
- Decoupling is Key: Freezing the weights of the bias-correction network (Stage 1) before training the classifier (Stage 2) forces the ResNet to rely on true physiological changes in tissue composition rather than scanner-specific intensity artifacts.
- Hadamard Efficiency: Unlike conventional convolutional layers, Hadamard-based layers operate in the transform domain, replacing expensive multiplications with sign-based additions, making the preprocessing step highly efficient.
- 2.5D Context Matters: Stacking adjacent slices ($N \pm 1$) provides vital inter-slice spatial consistency, significantly reducing false positives compared to purely 2D patch processing.
Key Results
The HBR-Net-18 framework demonstrated substantial performance gains over traditional radiomics and standard baseline CNNs trained on raw data:
Patch-Level Performance
- HBR-Net-18: 93.8% Accuracy | 94.4% Sensitivity | 93.3% Specificity
- Standard ResNet-18: 62.5% Accuracy | 97.9% Sensitivity | 60.5% Specificity
- Radiomics Baseline: 80.0% Accuracy | 91.0% Sensitivity | 75.0% Specificity
Patient-Level Performance
- HBR-Net-18: 83.7% Accuracy | 88.5% Sensitivity | 78.5% Specificity
- Radiomics Baseline: 60.0% Accuracy | 100.0% Sensitivity | 20.0% Specificity
The ablation study confirmed that removing the bias correction stage dropped patch-level accuracy from 93.8% to 89.1%, while removing the adjacent slice context (2D Only) dropped it to 88.2%.
Limitations and Future Work
The primary limitation of this study is its relatively small dataset (86 patients), which restricts the authors from exploring more complex architectures like Vision Transformers (ViTs) or hybrid models.
Additionally, the current patch-sampling strategy excluded benign tissue mimics (such as Benign Prostatic Hyperplasia [BPH] or inflammation) from cancerous patients during training. The authors note that future work will focus on incorporating these benign mimics into the training data to further improve localization accuracy and clinical robustness.
Summary updated