nnsam2: No-New SAM2 Segmentation Framework

• The paper introduces nnsam2, a pipeline that leverages a frozen SAM2 model and three sequential nnU-Net refinements to generate and refine pseudo-labels with minimal supervision.
• nnsam2 is a few-shot framework for multi-modality segmentation of lumbar paraspinal muscles, delivering high DSCs (up to 0.96 on MRI and 0.93 on CT) using one annotated slice per dataset.
• The approach employs sequential pseudo-label confidence scoring and strict DSC/area thresholds to enforce anatomical consistency and provide reliable quantitative measurements like muscle volume and CT attenuation.

No-New SAM2, abbreviated nnsam2, is a few-shot framework for multi-modality segmentation of lumbar paraspinal muscles that uses SAM2 without fine-tuning or architectural modification and then refines the resulting pseudo-labels with three sequential, independent nnU-Net models (Zhang et al., 7 Oct 2025). It was introduced for segmentation of the combined multifidus (MF) and lumbar erector spinae (LES) in the L4–L5 region, operationalized as slices spanning L3/L4 to L5/S1 disc levels, across multi-sequence MRI and multi-protocol CT (Zhang et al., 7 Oct 2025). In the reported minimal-supervision regime, one annotated slice per dataset was used to seed SAM2, and the framework was evaluated not only as a segmentation system but also as a pipeline for muscle volume, Dixon fat ratio, and CT attenuation measurement (Zhang et al., 7 Oct 2025).

1. Task definition and supervision regime

nnsam2 was developed for lumbar paraspinal muscle (LPM) analysis in heterogeneous cross-domain imaging cohorts. The study used 1,219 scans (19,439 slices) from 762 participants across six datasets, including 943 MRI scans from 486 participants and 276 CT scans from 276 participants (Zhang et al., 7 Oct 2025). The supervision budget was deliberately minimal: six slices, one per dataset, served as labeled examples, and the remaining 19,433 slices were used for testing (Zhang et al., 7 Oct 2025).

The target structures were the left and right lumbar paraspinal muscles in axial MRI or CT around the lower lumbar region. MF and LES were segmented jointly as one class per side, because they cannot be reliably distinguished on CT; the same convention was then used on MRI for consistency (Zhang et al., 7 Oct 2025). This design tied the method directly to downstream composition analysis rather than to class-rich anatomical parsing.

The six source datasets spanned markedly different acquisition settings. MRI cohorts included AFL T2-weighted MRI, back-pain T1-weighted MRI, back-pain T2-weighted MRI, and AGBRESA Dixon MRI; CT cohorts included TotalSegmentator CT and WORD CT (Zhang et al., 7 Oct 2025). This heterogeneity is central to the method’s intended role as a few-shot, cross-dataset, cross-modality framework rather than a sequence-specific or site-specific segmenter.

2. Meaning of “No-New SAM2”

In nnsam2, “No-New SAM2” means that SAM2 itself is not fine-tuned and not architecturally modified (Zhang et al., 7 Oct 2025). The contribution is therefore not a new SAM2 checkpoint, adapter stack, or prompt encoder, but a pipeline that uses frozen SAM2 for prompt-based pseudo-label generation and nnU-Net for 3D contextual refinement (Zhang et al., 7 Oct 2025). A plausible implication is that the phrase “No-New” applies specifically to the SAM2 component rather than to the absence of all additional learning stages.

This places nnsam2 at a conservative point within the broader SAM2 adaptation literature. Other SAM2-based systems add trainable adapters while freezing only part of the backbone, as in SAM2-Adapter (Chen et al., 2024), repackage the SAM2 encoder inside a U-shaped segmentation model, as in SAM2-UNet (Xiong et al., 2024), wrap SAM2 in a training-free detector-guided tracking system, as in Seg2Track-SAM2 (Mendonça et al., 15 Sep 2025), or use SAM2 directly in a zero-shot pseudo-video medical pipeline, as in 3D knee MRI segmentation with SAM2 (Yu et al., 2024). Relative to those variants, nnsam2 preserves the original SAM2 model more strictly while moving the main learning burden into a downstream refinement cascade.

3. Pseudo-label generation with frozen SAM2

The first stage of nnsam2 is single-slice SAM2 pseudo-label generation (Zhang et al., 7 Oct 2025). Before SAM2 inference, MRI intensities were clipped to the 0.5th–99.5th percentiles, CT intensities were clipped to $[-30,150]$ HU, CT volumes were downsampled along the head-foot axis by 3× for TotalSegmentator CT and 2× for WORD CT, and all MR and CT images were rescaled to $[0,255]$, converted to JPEG, and resized to $256\times256$ (Zhang et al., 7 Oct 2025).

For each dataset, a reference volume was selected as the volume with the smallest average feature distance across the L3/L4, L4/L5, and L5/S1 disc slices using DINOv2 features (Zhang et al., 7 Oct 2025). A single manually annotated slice from that reference volume—the top slice—served as the SAM2 prompt seed (Zhang et al., 7 Oct 2025). SAM2 was then applied through an interleaved sequence that alternated slices from the reference volume and each target inference volume, proceeding from superior to inferior slices across the lumbar region (Zhang et al., 7 Oct 2025).

This design allowed SAM2’s sequential propagation mechanism to transfer one prompted annotation into pseudo-labels for the remainder of a dataset (Zhang et al., 7 Oct 2025). For each predicted mask, SAM2 also produced an IoU score, which was treated as a confidence estimate and later used for pseudo-label selection (Zhang et al., 7 Oct 2025). The output of this stage was therefore a confidence-ranked set of SAM2 pseudo-labels rather than a final segmentation.

4. Three-stage nnU-Net refinement cascade

The second stage of nnsam2 refines SAM2 pseudo-labels through three sequential, independent nnU-Net models (Zhang et al., 7 Oct 2025). The purpose of this cascade is to suppress pseudo-label noise while introducing 3D anatomical consistency that SAM2 alone does not impose in this application (Zhang et al., 7 Oct 2025).

In Refinement Step 1, the method retained the top 10% of SAM2 masks ranked by IoU at the dataset level and the top 2% at the slice level, then pooled the selected pseudo-labels from all datasets to train nnU-Net 1 (Zhang et al., 7 Oct 2025). The reported rationale was empirical: 5% gave too few masks, whereas 20% introduced too much noise; the slice-level 2% preserved anatomically difficult slices that might otherwise be lost (Zhang et al., 7 Oct 2025).

In Refinement Step 2, nnU-Net 1 predictions were kept only if their mean DSC with the SAM2 prediction was $>0.90$ and their cross-sectional area was no larger than $1.5\times$ that of the superior slice (Zhang et al., 7 Oct 2025). From these validated predictions, the top 10% were used to train nnU-Net 2 (Zhang et al., 7 Oct 2025). In Refinement Step 3, nnU-Net 2 predictions were compared against nnU-Net 1 predictions, retained only if DSC $>0.90$ and area was $\le 1.25\times$ the superior slice area, and the top 20% were used to train nnU-Net 3, which served as the final predictor (Zhang et al., 7 Oct 2025). The threshold logic was again empirical: 0.85 admitted noisy masks, 0.95 excluded too many valid masks, $2.0\times$ area growth was too loose, and $1.25\times$ was too strict too early (Zhang et al., 7 Oct 2025).

The nnU-Net implementation used nnU-Net version 1.7.1, trained for 1,000 epochs with initial learning rate 0.01 and hybrid Dice-cross-entropy loss (Zhang et al., 7 Oct 2025). Mirroring was disabled to avoid anatomical mislabeling, and postprocessing kept only the largest connected region per muscle class (Zhang et al., 7 Oct 2025). This suggests that nnsam2 relies on progressively cleaner supervision and explicit anatomical smoothness constraints rather than a single pseudo-labeling pass.

5. Datasets, baselines, and segmentation accuracy

nnsam2 was evaluated against vanilla SAM2, MedSAM2, Medical SAM2, FAMNet, and TotalSegmentator (Zhang et al., 7 Oct 2025). The abstract reports that nnsam2 outperformed vanilla SAM2, its medical variants, TotalSegmentator, and the leading few-shot method, achieving DSCs of 0.94–0.96 on MR images and 0.92–0.93 on CT (Zhang et al., 7 Oct 2025).

The per-dataset numbers reported in the study were consistent with that summary. On MRI, nnsam2 reached 0.95/0.95 left/right DSC on AFL T2W, 0.96/0.96 on back-pain T2W, 0.96/0.96 on back-pain T1W, and 0.94/0.94 on AGBRESA Dixon (Zhang et al., 7 Oct 2025). On CT, it reached 0.92/0.93 on TotalSegmentator CT and 0.92/0.92 on WORD CT (Zhang et al., 7 Oct 2025). The paper summary further notes that improvements over vanilla SAM2 were generally around 0.02–0.03 DSC, whereas improvements over FAMNet were much larger and often exceeded 0.20 (Zhang et al., 7 Oct 2025).

The baseline behavior is also informative. Vanilla SAM2 was already a strong pseudo-label generator in this setting, with DSCs around 0.90–0.94 across the six datasets, but it did not match the refined nnsam2 outputs (Zhang et al., 7 Oct 2025). Medical SAM2 and MedSAM2 were less stable across the mixed cohorts, and TotalSegmentator—used off the shelf rather than fine-tuned—did not reach nnsam2 performance on the target muscles (Zhang et al., 7 Oct 2025). This suggests that the main gain of nnsam2 came from converting sparse prompt-based pseudo-labels into a pooled, denoised, anatomically regularized volumetric training signal.

6. Composition analysis and statistical agreement

A defining characteristic of nnsam2 is that it was evaluated not only by segmentation overlap but also by the statistical comparability of derived muscle measurements (Zhang et al., 7 Oct 2025). The reported endpoints were muscle volume from MRI and CT, fat ratio from Dixon MRI, and CT attenuation in Hounsfield units (Zhang et al., 7 Oct 2025). According to the abstract, automated and expert measurements were statistically equivalent for muscle volume (MRI/CT), CT attenuation, and Dixon fat ratio, with TOST $P < 0.05$ and ICCs ranging from 0.86 to 1.00 (Zhang et al., 7 Oct 2025).

The statistical framework combined two one-sided tests (TOST), intraclass correlation coefficients (ICC), and model-based Bland–Altman analysis (Zhang et al., 7 Oct 2025). The paper gives the ICC as

$[0,255]$0

where $[0,255]$1 is variance across individuals and $[0,255]$2 is residual or within-subject variance (Zhang et al., 7 Oct 2025). For Bland–Altman analysis using linear mixed-effects models, the limits of agreement were

$[0,255]$3

Equivalence in TOST was concluded when the 90% confidence interval for the estimated method effect lay entirely within the predefined equivalence bounds, with equivalence accepted if $[0,255]$4 (Zhang et al., 7 Oct 2025).

The reported measurement agreement was strongest on MRI, especially in the back-pain cohorts, where ICCs for volume reached 1.00 on both sides for both T1W and T2W MRI (Zhang et al., 7 Oct 2025). On AGBRESA Dixon MRI, fat-ratio ICCs were 0.97 and 0.96 for left and right sides, respectively (Zhang et al., 7 Oct 2025). CT remained slightly harder, but volume and attenuation still reached high agreement, with CT attenuation ICCs up to 0.99 on TotalSegmentator CT (Zhang et al., 7 Oct 2025). This measurement layer is central to the clinical framing of nnsam2: segmentation quality was treated as a means to reproducible quantitative phenotype extraction rather than as the sole endpoint.

7. Limitations and position within SAM2 adaptation research

The study identifies several limitations. Manual references were produced by a trained annotator plus one clinical professor reviewer, rather than multiple independent expert raters (Zhang et al., 7 Oct 2025). MF and LES were grouped into a single class, which constrains finer-grained anatomical analysis (Zhang et al., 7 Oct 2025). The framework also depends on nnU-Net refinement, and the paper notes that if data are insufficient, one may need to use SAM2 directly without nnU-Net enhancement (Zhang et al., 7 Oct 2025). Reported failure cases occurred in scans with poor image resolution or indistinct muscle boundaries (Zhang et al., 7 Oct 2025).

Within the broader SAM2 ecosystem, nnsam2 occupies a distinct methodological niche. Some medical SAM2 systems reinterpret volumes as videos and run zero-shot inference with no additional training, as in 3D knee MRI segmentation with SAM2 (Yu et al., 2024). Others remove manual prompting through support sets and LoRA, as in SAM2-SGP (Xing et al., 24 Jun 2025), or introduce explicit auxiliary modules for volumetric continuity and boundary supervision, as in SAM2-3dMed (Yang et al., 10 Oct 2025) and UniUltra (Li et al., 19 Nov 2025). By contrast, nnsam2 keeps SAM2 unchanged and shifts adaptation into a downstream pseudo-label refinement regime (Zhang et al., 7 Oct 2025). This suggests that nnsam2 is best understood not as a new SAM2 model, but as a few-shot bootstrap strategy that exploits frozen SAM2 generalization and then regularizes it with sequential nnU-Net refinement.

The study also released open code and data, including manual annotations for 826 MRI scans (13,633 slices) and 276 CT scans (3,832 slices), via the project repository reported in the paper (Zhang et al., 7 Oct 2025). That release reinforces the framework’s emphasis on annotation efficiency, cross-dataset reproducibility, and multi-modality evaluation under a supervision budget of only one labeled slice per dataset.