⚡ electrical engineering

Reinforcing the Weakest Links: Modernizing SIENA with Targeted Deep Learning Integration

This study demonstrates that strategically integrating deep learning tools like SynthStrip and SynthSeg into the established SIENA pipeline significantly enhances the accuracy, robustness, and efficiency of brain atrophy estimation by addressing its most vulnerable image processing steps while preserving the framework's interpretability.

Original authors: Riccardo Raciti, Lemuel Puglisi, Francesco Guarnera, Daniele Ravì, Sebastiano Battiato

Published 2026-03-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Riccardo Raciti, Lemuel Puglisi, Francesco Guarnera, Daniele Ravì, Sebastiano Battiato

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 your brain is a house. Over time, especially as we age or face diseases like Alzheimer's or Parkinson's, the walls of this house slowly shrink, and the rooms get smaller. This shrinking is called brain atrophy.

Doctors need to measure exactly how much the house is shrinking to see if a new medicine is working. They use MRI scans (like taking photos of the house) and a special software tool called SIENA to calculate the shrinkage.

For years, SIENA has been the "gold standard" tool for this job. But, like any old tool, it has some weak spots. Specifically, the part of the tool that tries to separate the "brain" from the "skull" (like peeling an orange without tearing the fruit) is a bit clumsy. If it makes a small mistake here, that error ripples through the rest of the process, leading to inaccurate measurements of how much the house has shrunk.

The Problem: The "Clumsy Peeler"

The original SIENA tool uses an old-school method to peel away the skull. Think of it like a person trying to peel an orange in the dark using a dull knife. Sometimes they cut too deep (removing part of the brain) or not deep enough (leaving part of the skull on). When this happens, the final calculation of the house's shrinkage is wrong.

The Solution: "Smart Glasses" for the Software

The authors of this paper asked: What if we keep the rest of the trusted SIENA tool, but swap out that clumsy peeling step with a modern, AI-powered "smart peeler"?

They didn't throw away the whole tool (which would be risky and confusing for doctors). Instead, they used a strategy called "Reinforcing the Weakest Links." They took two specific, modern AI tools:

SynthStrip: A super-smart AI that is incredibly good at peeling the skull off the brain.
SynthSeg: An AI that is excellent at identifying the different types of tissue inside the brain.

They plugged these AI tools into the old SIENA machine and tested three new versions:

Version A: Old peeler, new tissue identifier.
Version B: New peeler, old tissue identifier.
Version C: New peeler, new tissue identifier.

The Results: A Much Sharper Picture

They tested these new versions on thousands of patients with Alzheimer's and Parkinson's. Here is what they found, using simple analogies:

1. The "New Peeler" Changed Everything
The biggest improvement came from swapping out the skull-peeling step (SynthStrip).

Analogy: Imagine trying to measure the size of a shrinking balloon. If your ruler is crooked because you didn't peel the wrapper off the balloon correctly, your measurement is useless. Once they used the "smart peeler," the measurements became incredibly accurate.
Real-world impact: The new version was much better at connecting the brain shrinkage to how sick the patient actually felt. It was like upgrading from a blurry black-and-white photo to a crystal-clear HD video. The link between the MRI data and the patient's memory loss became much stronger.

2. No More "Which Way is Up?" Confusion
In the old tool, if you took the "before" and "after" photos and swapped their order, the computer sometimes gave a slightly different answer. That's like a scale that says you weigh 150 lbs one minute and 152 lbs the next just because you stepped on it backward.

The Fix: The new AI-powered versions were rock-solid. They gave the exact same answer (just with the sign flipped) regardless of the order of the photos. They reduced this "wobble" by up to 99%. It's like switching from a shaky hand-held camera to a tripod-mounted one.

3. Speeding Up the Process
The old tool took about 31 minutes to process one patient's scans on a standard computer.

The Fix: When they used a powerful graphics card (GPU) with the new AI tools, the time dropped to about 16 minutes. It's like switching from a bicycle to a sports car. The old tool was still fast enough on a regular computer, but the new one is a rocket ship when given the right fuel.

The Big Takeaway

The most important lesson from this paper isn't that we need to throw away old, trusted medical tools. Instead, it's that we can modernize them by upgrading their weakest parts.

Think of it like restoring a classic car. You don't replace the whole engine or the chassis because they are reliable and well-understood. Instead, you swap out the old, rusty carburetor for a modern, fuel-injected one. The car still looks and feels like the classic model, but it runs smoother, faster, and more reliably.

By doing this, the researchers made a trusted medical tool more accurate, more consistent, and faster, helping doctors better track diseases like Alzheimer's and Parkinson's without losing the trust they have in the original method.

1. Problem Statement

Longitudinal Brain Atrophy Estimation: Quantifying brain volume loss (atrophy) over time is critical for monitoring neurodegenerative diseases like Alzheimer's (AD) and Parkinson's (PD). The standard metric is Percentage Brain Volume Change (PBVC) derived from paired MRI scans.

Limitations of Current Methods: The most established tool for PBVC estimation is SIENA (part of the FSL toolbox). While robust, SIENA relies on classical image processing algorithms:

BET2: For skull stripping (brain extraction).
FAST: For tissue segmentation (White Matter, Gray Matter, CSF).

The Core Issue: These classical algorithms rely on intensity-based heuristics and deformable surface models. They are prone to failure in the presence of severe neurodegeneration, anatomical abnormalities, or imaging artifacts. Crucially, errors in the initial skull stripping step propagate through the pipeline, causing cascading failures in registration and segmentation, which ultimately biases PBVC estimates. While end-to-end Deep Learning (DL) models exist, they often lack the interpretability and transparency of established clinical pipelines like SIENA.

2. Methodology

The authors propose a modular modernization strategy: replacing specific "weak links" in the SIENA pipeline with state-of-the-art Deep Learning models while preserving the core SIENA framework (registration, boundary shift estimation, and PBVC calculation).

Deep Learning Components Integrated:

SynthStrip: A domain-randomized DL model for brain extraction. It generates synthetic training data with varied contrast, noise, and artifacts to ensure high generalizability.
SynthSeg: A domain-randomized DL model for semantic tissue segmentation.

Pipeline Variants Evaluated:
The study tested four configurations on longitudinal T1-weighted MRI data:

SIENA Vanilla (Baseline): BET2 (skull stripping) + FAST (segmentation).
SIENA-SS: SynthStrip (skull stripping) + FAST (segmentation).
- Note: Since SynthStrip does not output a skull mask, the authors implemented a post-processing step to derive an approximate skull mask from the SynthStrip brain mask using Gaussian smoothing and ray-casting to detect the inner skull boundary.
SIENA-SEG: BET2 (skull stripping) + SynthSeg (segmentation).
- Note: SynthSeg outputs anatomical labels; these were aggregated into the three tissue classes (WM, GM, CSF) required by SIENA.
SIENA-SS-SEG: SynthStrip + SynthSeg (Fully integrated).

Datasets:

ADNI (Alzheimer's Disease Neuroimaging Initiative): 1,006 participants.
PPMI (Parkinson's Progression Markers Initiative): 310 participants.

Evaluation Metrics:

Clinical Sensitivity: Correlation between estimated PBVC and longitudinal changes in clinical scores (e.g., MMSE, ADAS-13, CDR-SB, MoCA) and structural measures (GM volume, BPF).
Scan-Order Consistency: Measuring the Mean Forward-Reverse Residual (MFRR). Ideally, $PBVC(A, B) = -PBVC(B, A)$. Deviation from this indicates pipeline instability.
Computational Efficiency: End-to-end runtime on CPU and GPU.

3. Key Contributions

Modular Integration: Demonstrated that Deep Learning can be integrated into legacy clinical pipelines without discarding their interpretability.
Identification of the "Weakest Link": Systematically proved that the skull stripping step is the primary source of error and bias in the SIENA pipeline, more so than tissue segmentation.
Robustness Improvement: Showed that replacing classical algorithms with domain-randomized DL models significantly improves the stability of atrophy estimates against scan-order reversal.
Open Source: Released the implementation code publicly to facilitate reproducibility.

4. Key Results

A. Correlation with Disease Progression (Clinical Sensitivity)

ADNI Cohort: Replacing skull stripping with SynthStrip (SIENA-SS) yielded the most significant improvements.
- Correlation with $\Delta$ CDR-SB improved from $r = -0.258$ (Vanilla) to $r = -0.608$ (SIENA-SS).
- Correlation with $\Delta$ ADAS-13 improved from $r = -0.254$ to $r = -0.524$ .
- Replacing only segmentation (SIENA-SEG) showed modest or non-significant gains compared to the baseline.
PPMI Cohort: Trends were similar (shifts toward expected negative correlations), though effect sizes were smaller and did not reach statistical significance after multiple-comparison correction, likely due to the smaller sample size and different disease pathology.

B. Scan-Order Consistency (Robustness)

This was the most dramatic finding. The standard SIENA pipeline is sensitive to the order of input scans (baseline vs. follow-up).

ADNI: The fully integrated pipeline (SIENA-SS-SEG) reduced the error (MFRR) by 87.8% (from 0.379% to 0.046%).
PPMI: The reduction was even more profound, with SIENA-SS and SIENA-SS-SEG reducing the error by 99.0% and 99.1% respectively (MFRR dropped to 0.002%).
Conclusion: The DL-based skull stripping (SynthStrip) is the primary driver of this robustness, as it provides more anatomically consistent brain masks regardless of scan order.

C. Execution Time

CPU: The modified pipelines maintained runtimes comparable to SIENA Vanilla (approx. 31 minutes per pair).
GPU: Significant speedups were achieved when using GPU acceleration for the segmentation step.
- SIENA-SEG and SIENA-SS-SEG reduced runtime by up to 46% (from ~31 mins to ~16-19 mins) compared to the CPU baseline.
- Replacing skull stripping (SynthStrip) offered only modest speedups on GPU as BET2 is already fast.

5. Significance and Conclusion

The paper argues that targeted modernization is a superior strategy to replacing entire clinical frameworks. By reinforcing the weakest links (specifically skull stripping) with robust, domain-randomized Deep Learning models, the authors achieved:

Higher Clinical Relevance: PBVC estimates correlate much more strongly with actual disease progression.
Greater Reliability: The pipeline becomes nearly invariant to the order of scan acquisition, a critical requirement for longitudinal studies.
Scalability: GPU acceleration allows for faster processing without sacrificing the interpretability of the SIENA framework.

Broader Impact: This work validates a "modular" approach to AI in medicine, where AI is used to upgrade specific components of trusted, legacy tools rather than replacing them entirely. This preserves clinical trust and interpretability while leveraging the superior generalization capabilities of modern Deep Learning.