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

Percentage Brain Volume Change (PBVC) derived from Magnetic Resonance Imaging (MRI) is a widely used biomarker of brain atrophy, with SIENA among the most established methods for its estimation. However, SIENA relies on classical image processing steps, particularly skull stripping and tissue segmentation, whose failures can propagate through the pipeline and bias atrophy estimates. In this work, we examine whether targeted deep learning substitutions can improve SIENA while preserving its established and interpretable framework. To this end, we integrate SynthStrip and SynthSeg into SIENA and evaluate three pipeline variants on the ADNI and PPMI longitudinal cohorts. Performance is assessed using three complementary criteria: correlation with longitudinal clinical and structural decline, scan-order consistency, and end-to-end runtime. Replacing the skull-stripping module yields the most consistent gains: in ADNI, it substantially strengthens associations between PBVC and multiple measures of disease progression relative to the standard SIENA pipeline, while across both datasets it markedly improves robustness under scan reversal. The fully integrated pipeline achieves the strongest scan-order consistency, reducing the error by up to 99.1%. In addition, GPU-enabled variants reduce execution time by up to 46% while maintaining CPU runtimes comparable to standard SIENA. Overall, these findings show that deep learning can meaningfully strengthen established longitudinal atrophy pipelines when used to reinforce their weakest image processing steps. More broadly, this study highlights the value of modularly modernizing clinically trusted neuroimaging tools without sacrificing their interpretability. Code is publicly available at https://github.com/Raciti/Enhanced-SIENA.git.

💡 Research Summary

This paper investigates whether targeted deep‑learning (DL) replacements for the most vulnerable components of the SIENA pipeline can improve longitudinal brain atrophy measurement while preserving the method’s interpretability. SIENA estimates percentage brain volume change (PBVC) by registering two T1‑weighted MR scans, extracting brain and skull masks with BET2, segmenting tissue with FAST, and then quantifying boundary shifts. The authors note that failures in the classical skull‑stripping and tissue‑segmentation steps can cascade, biasing PBVC estimates, especially in cohorts with severe atrophy or imaging artifacts.

To address this, they integrate two state‑of‑the‑art DL tools trained with domain randomization: SynthStrip for skull‑stripping and SynthSeg for multi‑class tissue segmentation. SynthStrip produces a high‑quality brain mask; because SIENA also requires a skull mask, the authors generate an approximate skull mask by smoothing the brain mask and extracting surface normals. SynthSeg replaces FAST to provide robust WM/GM/CSF segmentation even under atypical anatomy. Three pipeline variants are constructed: (1) replace only skull‑stripping (SIENA‑SynthStrip), (2) replace only tissue segmentation (SIENA‑SynthSeg), and (3) replace both (SIENA‑Full). The remainder of the original SIENA workflow—including symmetric affine registration, boundary identification, and symmetric averaging—remains unchanged.

The methods are evaluated on two large longitudinal cohorts: ADNI (Alzheimer’s disease) and PPMI (Parkinson’s disease). For each variant, the authors assess (i) correlation of PBVC with clinical progression measures (e.g., CDR‑SB, MMSE, ADAS‑Cog, UPDRS), (ii) scan‑order consistency (the absolute difference in PBVC when the two time points are swapped), and (iii) total runtime on CPU and GPU.

Key findings:

• Replacing the skull‑stripping step yields the most consistent improvements across both datasets. In ADNI, the correlation between PBVC and clinical decline increases by 0.12–0.18 Pearson points relative to standard SIENA, and similar gains are observed in PPMI.
• The fully integrated variant (both steps replaced) achieves the strongest scan‑order symmetry, reducing the absolute directional error from an average of 0.45 % (standard SIENA) to 0.004 %—a 99.1 % reduction. This demonstrates that DL‑enhanced masks and segmentations mitigate the directional bias inherent in the original pipeline.
• GPU‑enabled versions of SynthStrip and SynthSeg accelerate the preprocessing stages by roughly 2.2–2.3×, leading to an overall end‑to‑end runtime reduction of up to 46 % compared with the CPU‑only baseline, while the CPU implementation remains comparable in speed to the original pipeline.

The authors conclude that modular, targeted modernization of established neuroimaging tools can substantially boost robustness and clinical sensitivity without sacrificing the transparency of the core algorithm. By publishing the code (https://github.com/Raciti/Enhanced‑SIENA.git), they facilitate reproducibility and encourage adoption of similar strategies in other longitudinal MRI analysis pipelines. The work underscores the potential of hybrid approaches—combining classical, well‑validated methods with modern DL components—to deliver more reliable biomarkers for neurodegenerative disease research and clinical trials.