💻 computer science

Hybrid Multi-Dimensional MRI Prostate Cancer Detection via Hadamard Network-Based Bias Correction and Residual Networks

This paper proposes HBR-Net-18, a two-stage AI framework that combines a Hadamard U-Net for bias correction and a ResNet-18 for patch-level classification to achieve robust, automated prostate cancer detection using hybrid multi-dimensional MRI data.

Original authors: Emadeldeen Hamdan, Gorkem Durak, Muhammed Enes Tasci, Abel Lorente Campos, Aritrick Chatterjee, Roger Engelmann, Gregory Karczma, Aytekin Oto, Ahmet Enis Cetin, Ulas Bagci

Published 2026-04-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Emadeldeen Hamdan, Gorkem Durak, Muhammed Enes Tasci, Abel Lorente Campos, Aritrick Chatterjee, Roger Engelmann, Gregory Karczma, Aytekin Oto, Ahmet Enis Cetin, Ulas Bagci

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to find a hidden treasure (prostate cancer) inside a very complex, foggy landscape (the human body) using a special kind of map (MRI).

This paper presents a new, super-smart "detective team" called HBR-Net-18 that helps doctors find that treasure much more accurately than before. Here is how it works, broken down into simple steps with some fun analogies.

1. The Problem: The "Foggy Map"

Doctors use a special type of MRI called HM-MRI to look at the prostate. Think of this MRI not just as a picture, but as a set of six different colored maps that show the chemical makeup of the tissue (like how much water is in a cell or how fast it moves).

However, these maps often have a problem: Bias Fields.

The Analogy: Imagine looking at a beautiful landscape through a dirty, uneven window. Some parts of the view look bright, some look dark, and some look blurry, not because the landscape changed, but because the window is dirty.
The Issue: In MRI, this "dirt" is called a "bias field." It makes healthy tissue look like cancer, or cancer look healthy, confusing the computer (and the doctor).

2. The Solution: A Two-Stage Detective Team

The authors built a two-step AI system to solve this. Think of it as a team with two specialized agents.

Stage 1: The "Window Cleaner" (The Hadamard-Bias Network)

Before the detective looks for the treasure, someone has to clean the window.

What it does: This part of the AI uses a special mathematical tool (called a Hadamard Transform) to act like a high-tech window cleaner. It scans the six maps and smooths out the "dirt" (the uneven brightness).
The Magic Trick: It uses a technique called a "U-Net" (which looks like a U-shape in its design) to figure out exactly where the "dirt" is and remove it, leaving a perfectly clear, consistent map.
Why it matters: Once the window is clean, the next detective doesn't get tricked by shadows or bright spots. They only look at the actual shape of the land.

Stage 2: The "Patch Detective" (The ResNet-18)

Now that the maps are clean, the second detective gets to work.

The Strategy: Instead of looking at the whole map at once (which is too big and confusing), this detective cuts the map into tiny 11-by-11 puzzle pieces (patches).
The 3D Vision: Here is the clever part. The detective doesn't just look at one puzzle piece in isolation. They look at the piece, the one above it, and the one below it.
- The Analogy: Imagine trying to guess what a character in a movie is doing. If you only look at one frozen frame, it's hard. But if you look at the frame before, the current one, and the one after, you understand the story much better. This AI does the same thing with the 3D slices of the MRI.
The Decision: It examines these tiny, 3D puzzle pieces and decides: "Is this piece cancer?" or "Is this healthy?"

3. The Results: Why This Team Wins

The researchers tested their new team against old methods (like standard radiology and older AI models).

The Old Way: The old methods were like trying to find a needle in a haystack while wearing sunglasses. They were okay, but they missed a lot of needles or thought the hay was a needle.
The New Way (HBR-Net-18): Because they cleaned the window first and looked at the 3D context, their team was incredibly accurate.
- They caught 94.4% of the actual cancers (Sensitivity).
- They correctly identified 93.3% of the healthy tissue as healthy (Specificity).
- This is a huge jump compared to older AI models, which often got confused by the "dirty window" and made mistakes.

4. The Bottom Line

This paper introduces a smarter way to use AI for prostate cancer.

First, it fixes the "bad lighting" in the MRI scans so the data is fair and consistent.
Second, it looks at the data in small, connected chunks (like a 3D puzzle) to spot the cancer.

The Takeaway: By cleaning up the data first and then looking at the "big picture" through small, connected windows, this new AI system acts like a super-powered magnifying glass, helping doctors find prostate cancer earlier and with much more confidence. It's a step forward toward making cancer diagnosis faster, more automatic, and less prone to human error.

1. Problem Statement

Prostate cancer (PCa) is a leading cause of cancer-related mortality in men. While Hybrid Multi-Dimensional MRI (HM-MRI) offers superior soft-tissue contrast and quantitative tissue characterization (modeling stroma, epithelium, and lumen), its clinical utility is hindered by:

Bias Field Inhomogeneities: Low-frequency intensity variations across MRI sequences that distort tissue appearance.
Signal Variability: Inconsistencies between different MRI acquisitions.
Limited Generalization: Standard Convolutional Neural Networks (CNNs) often struggle with inter-slice inconsistencies and may rely on spurious intensity cues rather than true pathological features.
Data Scarcity: The need for robust automated detection methods that do not rely on massive annotated datasets.

The authors aim to develop a robust, automated AI framework that explicitly corrects these intensity inhomogeneities before performing cancer classification, thereby improving diagnostic reliability.

2. Methodology: HBR-Net-18

The authors propose HBR-Net-18, a modular, two-stage AI framework designed to decouple bias correction from cancer classification.

Input Data

The system utilizes six quantitative tissue biomarker maps generated by a Physics-Informed Autoencoder (PIA) from HM-MRI data:

Epithelial volume fraction ( $v_{ep}$ )
Luminal water fraction ( $v_{lu}$ )
Epithelial diffusivity ( $d_{ep}$ )
Stromal diffusivity ( $d_{st}$ )
Epithelial T2 relaxation time ( $t_{2ep}$ )
Stromal T2 relaxation time ( $t_{2st}$ )

Stage 1: Bias-Aware Biomarker Correction (Hadamard U-Net)

Goal: Suppress low-frequency intensity inhomogeneities (bias fields) and noise while preserving diagnostically relevant spatial structures.
Architecture: A probabilistic Hadamard U-Net.
Mechanism:
- Utilizes the Walsh–Hadamard Transform (WHT), an orthogonal linear transform that decomposes signals into binary-valued functions.
- Replaces standard multiplications with sign-based additions in the transform domain, enabling energy-efficient computation.
- The network is trained to map raw biomarker maps ( $X$ ) to bias-corrected maps ( $\hat{X}$ ) by minimizing the Mean Squared Error (MSE) against reference maps generated by the N4ITK algorithm.
Key Constraint: Once trained, the parameters ( $\theta$ ) of this stage are frozen. This enforces a strict separation between bias normalization and cancer decision-making, preventing the classifier from "cheating" by learning acquisition-specific intensity artifacts.

Stage 2: Patch-Level Cancer Classification (ResNet-18)

Data Preparation:
- Corrected maps are confined to the prostate region (using expert masks).
- The volume is divided into overlapping 11 × 11 patches with a stride of 2.
- 3D Context: To improve spatial consistency, each target slice ( $N$ ) is augmented with adjacent slices ( $N-1, N, N+1$ ), creating a tensor of size $3 \times 6 \times 11 \times 11$ .
Architecture:
- Upsampling Module: A learnable convolutional projection expands the patch dimensions to $64 \times 32 \times 32$ .
- Classifier: A ResNet-18 network processes the upsampled features to output a probability ( $p_i \in [0,1]$ ) for cancer presence.
Training: The classifier is trained end-to-end using Focal Loss to address class imbalance between cancerous and non-cancerous patches.

3. Key Contributions

Decoupled Two-Stage Framework: The novel separation of bias correction (Stage 1) and classification (Stage 2) ensures the model learns robust physiological features rather than scanner artifacts.
Hadamard-Based Bias Correction: The integration of the Walsh–Hadamard Transform into a U-Net architecture provides an efficient, energy-saving method for correcting low-frequency intensity inhomogeneities in quantitative MRI.
3D Spatial Consistency: By incorporating adjacent slices ( $N \pm 1$ ) into the patch processing, the model captures inter-slice context, addressing the limitations of purely 2D analysis.
Physics-Informed Integration: The framework leverages PIA-derived biomarkers, which offer a deeper physiological interpretation of tissue microstructure compared to standard signal intensities.

4. Experimental Results

The study was conducted on data from 86 patients (34 cancer-free, 52 with cancer) using a 3T Philips MRI scanner. A 5-fold cross-validation strategy was employed.

Patch-Level Performance (Table I)

HBR-Net-18 significantly outperformed baselines:

Accuracy: 93.8% (vs. 80.0% for Radiomics, 62.5% for ResNet-18 baseline).
Sensitivity: 94.4%
Specificity: 93.3%
Improvement: It achieved a 17.25% accuracy gain over Radiomics and a 50.08% gain over a standard ResNet-18 trained on raw data.

Patient-Level Performance (Table II)

Accuracy: 83.7%
Specificity: 78.5% (Crucial for reducing false positives in healthy patients).
Sensitivity: 88.5%
Comparison: The model improved accuracy by an average of 21% over baseline models and showed a massive 292.50% improvement in specificity over Radiomics for cancer-free patients.

Ablation Study (Table III)

Removing Stage 1 (Bias Correction) dropped accuracy to 89.1%, proving the necessity of intensity normalization.
Removing 3D context (2D only) dropped accuracy to 88.2%, confirming the value of inter-slice information.
The 11 × 11 patch size was found to be the optimal trade-off between local detail and context.

5. Significance and Conclusion

The paper demonstrates that HBR-Net-18 is a highly effective tool for automated prostate cancer detection. By explicitly correcting bias fields before classification and leveraging 3D spatial context, the model achieves a superior balance of sensitivity and specificity compared to traditional radiomics and standard CNNs.

Clinical Impact:

Reduced False Positives: High specificity is critical for avoiding unnecessary biopsies in benign cases.
Robustness: The decoupled architecture makes the system more robust to acquisition variations, a common hurdle in multi-center clinical deployment.
Future Work: The authors acknowledge limitations regarding the small dataset size and the exclusion of benign mimics (e.g., BPH, inflammation) from the training data. Future iterations aim to include these conditions to further enhance localization accuracy and clinical generalizability.