Slice-wise quality assessment of high b-value breast DWI via deep learning-based artifact detection

This study demonstrates that a DenseNet121-based deep learning model effectively detects hyper- and hypointense intensity artifacts on high b-value (1500 s/mm²) breast diffusion-weighted MRI slices, achieving high AUROC scores and providing promising results for automated quality assessment.

The Gist

Imagine you are a detective trying to solve a crime inside a person's body using a special camera called an MRI. This camera takes pictures of the breast to find hidden tumors. To make the pictures clearer, the doctors use a special setting called "high b-value," which acts like a high-contrast filter to make healthy tissue disappear and make suspicious lumps stand out.

However, just like a camera lens can get smudged or a photo can have a weird flash, these MRI pictures often get "glitches." These glitches are called artifacts. Sometimes they look like bright white spots (hyperintense), and sometimes they look like dark black holes (hypointense).

The problem? These glitches can look exactly like tumors, or they can hide real tumors. If a doctor mistakes a glitch for a tumor, the patient gets scared for no reason. If they miss a real tumor because a glitch covered it, that's dangerous.

The Mission: The AI "Glitch Hunter"
The researchers in this paper wanted to build a robot detective (an Artificial Intelligence) that could look at these MRI slices and say, "Hey, that's just a glitch, not a tumor!" or "That's a real tumor, ignore the glitch."

Here is how they did it, broken down into simple steps:

1. Gathering the Evidence (The Dataset)

They collected over 11,000 individual slices (like pages in a book) from breast MRI scans. They didn't just look at the whole 3D picture; they looked at every single page one by one. This is important because a glitch might only appear on one page of the book, not the whole story.

2. Training the Robot (The School)

They taught three different types of "student robots" (AI models named DenseNet121, ResNet18, and SEResNet50) to spot these glitches.

• The Task: The robots had to learn two things:
  1. Is there a glitch? (Yes/No)
  2. How bad is the glitch? (Is it a tiny smudge or a massive blackout?)
• The Teachers: Real human doctors (radiologists) acted as the teachers, grading the pictures and telling the robots what was a glitch and what wasn't.

3. The Big Test (The Exam)

After the robots studied, they took a final exam on pictures they had never seen before.

• The Winner: One robot, DenseNet121, was the star student. It was incredibly good at spotting both the bright white glitches and the dark black glitches.
  • It was right about 92% of the time for bright glitches.
  • It was right about 94% of the time for dark glitches.
• The "Severe" Glitch Specialist: The robot was especially good at spotting the worst glitches (the ones that would ruin a diagnosis). It almost never confused a massive glitch with a clean picture. This is like a security guard who never misses a real intruder, even if they sometimes get a little jumpy about a shadow.

4. Drawing the Target (The Bounding Box)

To make sure the robot knew where the glitch was, the researchers used a special trick called Grad-CAM. Imagine the robot putting a glowing red box around the glitch to say, "Look here!"

• The Result: For the bright glitches, the box was usually pretty accurate (like a good bullseye). For the dark glitches, the box was a bit looser, sometimes including a little extra space, but it still found the general area.

Why Does This Matter?

Think of this AI as a quality control inspector for the MRI machine.

• Before: A technician takes a picture, and a doctor has to stare at it for a long time, wondering, "Is that a tumor or just a weird shadow?"
• After: The AI scans the picture instantly. If it sees a glitch, it flags it.
  • If the glitch is minor, the doctor can ignore it and keep looking.
  • If the glitch is severe, the AI can tell the technician, "Hey, this picture is ruined! Let's take it again immediately."

The Catch (Limitations)

The researchers were honest about the robot's flaws:

• It's a bit subjective: Sometimes even human doctors disagree on whether a glitch is "minor" or "moderate." If the teachers disagree, the student gets confused.
• It's a bit narrow: The robot was trained on pictures from just one hospital. It might get confused if it sees pictures from a different machine or a different country.
• It's not a tumor finder: The robot is only there to find the glitches. It doesn't find the cancer; it just clears the path so the human doctor can find the cancer better.

The Bottom Line

This paper shows that we can teach computers to spot the "noise" in medical pictures. By filtering out the static and the glitches, we help doctors see the truth more clearly, leading to fewer mistakes and less anxiety for patients. It's like cleaning a dirty window so you can finally see the view clearly.

Technical Summary

1. Problem Statement

Diffusion-weighted imaging (DWI) is increasingly integrated into breast MRI protocols to aid in lesion detection and characterization. However, high b-value DWI acquisitions (specifically $b = 1500 \text{ s/mm}^2$) are highly susceptible to intensity artifacts that can obscure pathology or mimic lesions. These artifacts generally fall into two categories:

• Hyperintense artifacts: Caused by skin folds, failed fat suppression, or coil flares.
• Hypointense artifacts: Caused by pulsation or susceptibility-related signal loss.

Existing methods, such as those by Kapsner et al., focused on detecting artifacts in Maximum Intensity Projections (MIPs). However, MIPs can obscure the specific slice location of an artifact, and artifacts originating from a small volume portion may distort the entire projection. Furthermore, previous work primarily addressed hyperintense artifacts. There is a need for a slice-wise approach that can detect and grade both hyper- and hypointense artifacts to improve quality control, prevent biased AI training, and guide technicians on re-scanning or protocol adjustments.

2. Methodology

Dataset and Preprocessing

• Data Source: A retrospective, IRB-approved single-center study using 3T MRI systems (Siemens Healthineers).
• Cohort: 1,383 routine breast MRI examinations (2022–mid-2023).
• Selection: 156 cases were pre-selected based on MIP analysis to ensure the presence of moderate-to-severe artifacts.
• Slice Generation: The 156 cases were converted into a slice-wise dataset. Slices were split into left and right breasts, resulting in 11,806 total slices.
• Preprocessing:
  • Images were rescaled to 0–255 and resized to $160 \times 128$.
  • Masking: Breast-only masks were generated from T1-weighted images (due to poor SNR in high b-value DWI) and applied to DWI slices to exclude background noise.
  • Labeling: A radiologist and a master's student annotated slices on a 6-point Likert scale (1: No artifact to 6: Ambiguous). Scores 1–2 were grouped as "No Artifact," and 3–5 as "Significant Artifact."
• Data Split: Stratified splitting at the case level (to prevent data leakage) resulted in:
  • Training: ~8,164 slices
  • Validation: ~1,800 slices
  • Holdout Test: ~1,800 slices

Deep Learning Architecture

• Models: Three Convolutional Neural Network (CNN) architectures were evaluated: DenseNet121, ResNet18, and SEResNet50.
• Tasks:
  1. Binary Classification: Distinguishing between "No/Minor Artifact" (Class 0) and "Significant Artifact" (Class 1).
  2. Multiclass Classification: Grading severity (Classes 1–5).
• Training Setup:
  • Framework: MONAI with PyTorch Lightning.
  • Hardware: NVIDIA RTX 2080 GPUs.
  • Optimization: Adam optimizer, Cross-Entropy loss, weighted random sampling to handle class imbalance.
  • Augmentation: Random rotation ($\pm 12^\circ$) and flipping.
  • Early Stopping: Based on validation loss (patience = 10).

Visualization and Evaluation

• Localization: Since ground truth bounding boxes were not available, Grad-CAM heatmaps were used to generate bounding boxes around the most influential regions for the model's prediction.
• Metrics:
  • Quantitative: AUROC, AUPRC, Accuracy, Precision, Recall, and Confusion Matrices.
  • Qualitative: A radiologist evaluated the bounding box precision on a 5-point Likert scale (1: No overlap to 5: Precise location).
  • Agreement: Cohen's Kappa coefficient was used to measure agreement between the model and human readers (Ground Truth, Validator 1, Validator 2).

3. Key Results

Model Performance (Binary Classification)

DenseNet121 outperformed ResNet18 and SEResNet50 across both artifact types on the holdout test set:

• Hyperintense Artifacts:
  • AUROC: 0.92
  • AUPRC: 0.77
  • Recall: 0.82
  • Bounding Box Mean Score: 3.33 ± 1.04
• Hypointense Artifacts:
  • AUROC: 0.94
  • AUPRC: 0.92
  • Recall: 0.91
  • Bounding Box Mean Score: 2.62 ± 0.81 (Lower than hyperintense, indicating harder localization).

Multiclass Classification

• Hyperintense: Weighted AUROC of 0.85. Notably, the model never misclassified severe artifacts (Class 5) as no artifact (Class 1) or minor artifact (Class 2).
• Hypointense: Weighted AUROC of 0.88.
• Reader Agreement: The model showed "fair" to "moderate" agreement with human readers (Cohen's Kappa), suggesting that while the model is robust, human labeling of subtle artifacts remains subjective.

4. Key Contributions

1. Dual-Artifact Detection: First study to simultaneously detect and grade both hyper- and hypointense artifacts in high b-value breast DWI on a slice-wise level, rather than relying on MIPs.
2. Architectural Benchmarking: Demonstrated that DenseNet121 is superior to ResNet18 and SEResNet50 for this specific task, likely due to its dense connectivity preserving low-level spatial features critical for subtle intensity changes.
3. Severity Grading: Successfully implemented a multiclass approach to distinguish between minor, moderate, and severe artifacts, which is crucial for clinical decision-making (e.g., deciding whether to rescan).
4. Explainability: Utilized Grad-CAM to generate bounding boxes, providing a visual explanation of where the model detected artifacts, despite the lack of precise ground-truth bounding boxes.

5. Significance and Limitations

Significance:

• Quality Control: The system can serve as an automated quality control tool in clinical workflows, flagging scans with significant artifacts for immediate review or re-acquisition.
• AI Robustness: By filtering out or flagging artifact-heavy slices, the reliability of downstream AI tasks (e.g., lesion segmentation, virtual contrast enhancement) is improved.
• Clinical Workflow: The ability to distinguish artifact severity helps technicians decide on mitigation strategies (e.g., changing fat saturation techniques) without unnecessary full re-scans.

Limitations:

• Localization Accuracy: The use of Grad-CAM for bounding boxes is an approximation; the scores were lower for hypointense artifacts, indicating the model sometimes struggles to pinpoint the exact artifact boundaries compared to dedicated object detection models (e.g., YOLO, Faster R-CNN).
• Subjectivity: Human labeling of artifact severity is subjective, leading to "label noise" that affects model training, particularly for intermediate severity classes.
• Generalizability: The dataset is single-center and lacks multi-center validation. The model was trained only on $b=1500 \text{ s/mm}^2$ and may not generalize to lower b-values without retraining.
• Clinical Impact: The study did not include cases where artifacts specifically obscured malignant lesions, so the direct impact on cancer detection rates remains to be proven.

Conclusion:
The study demonstrates that deep learning, specifically DenseNet121, is highly effective at detecting and grading artifacts in high b-value breast DWI. While localization requires refinement, the high AUROC scores suggest the model is ready for further validation as a clinical quality assurance tool.