Skeletal Muscle Area (SMA) in Medical Imaging

• Skeletal Muscle Area (SMA) is the cross-sectional measurement of muscle tissue from a single image plane, typically using the standardized L3 vertebral level.
• It is applied in monitoring muscle atrophy in ICU, evaluating sarcopenia, and tracking cancer cachexia, providing critical insights for clinical interventions.
• Automated computational pipelines using CT, MRI, and ultrasound enhance SMA extraction through precise segmentation and reproducible slice localization.

Searching arXiv for the provided SMA-related papers to ground the article in current literature. Skeletal Muscle Area (SMA) is the cross-sectional area of muscle tissue on a single image plane. In radiologic practice, recent work defines SMA as the cross-sectional area of all muscle tissue, measured in cm², on a single axial CT slice; related MRI and ultrasound workflows compute the same quantity from segmented muscle pixels and physical pixel area. SMA is used in longitudinal monitoring of muscle atrophy and recovery in intensive care, in sarcopenia assessment, and in identification and tracking of cancer cachexia. A major standardization convention is the mid-third lumbar vertebral level (L3), where prior studies cited in recent work report a high correlation ($r\approx0.95$) between muscle area and whole-body muscle volume, whereas longitudinal ultrasound studies emphasize retrieval of the same anatomical view across visits rather than a vertebral landmark (Ahmed et al., 19 Mar 2025, Kerdegari et al., 2023).

1. Definition, anatomical scope, and derived indices

In the CT-based oncology literature, SMA is defined as the cross-sectional area of all muscle tissue on a single axial CT slice, measured in cm² (Ahmed et al., 19 Mar 2025). In one CT sarcopenia workflow, manual annotation at L3 explicitly included the psoas, erector spinae, quadratus lumborum, and abdominal wall muscles, and the traced contour was rasterized into a binary mask whose foreground pixels corresponded to the expert’s SMA measurement (Bhardwaj et al., 24 Aug 2025). MRI and ultrasound studies use the same geometric principle: a muscle mask is estimated on a selected slice or view, and area is computed from the count of muscle pixels and the physical pixel area (Akella et al., 2024, Kerdegari et al., 2023).

The L3 convention is central in abdominal CT and MRI body-composition analysis. Recent work cites Shen et al. (2004) and Mourtzakis et al. (2008) for the observation that muscle area at the L3 level correlates highly with whole-body muscle volume, with $r\approx0.95$, and notes that L3 is readily identified on routine abdominal CT scans as a standardized, reproducible anatomical landmark (Ahmed et al., 19 Mar 2025). By contrast, longitudinal ultrasound of the rectus femoris in ICU patients requires repeated acquisition of the same anatomical “slice” over time, because the principal source of error is not only boundary delineation but also inconsistency in view matching across days and operators (Kerdegari et al., 2023).

A derived index used in oncology is the Skeletal Muscle Index (SMI), defined as

$$\mathrm{SMI} \;=\; \frac{\mathrm{SMA}}{\bigl(\text{height in m}\bigr)^2} \quad \bigl[\mathrm{cm}^2/\mathrm{m}^2\bigr].$$

BMI is used alongside SMI to adjust for overall body habitus in downstream predictive models (Ahmed et al., 19 Mar 2025). A common misconception is to treat SMA and SMI as interchangeable; the cited formulation distinguishes absolute cross-sectional muscle area from a height-normalized index.

2. Image acquisition and slice or view localization

Accurate SMA estimation begins with correct slice or view selection. In the SMAART-AI CT pipeline, all DICOM CT series are retrieved from PACS/VNA, axial series are identified via the ImageOrientationPatient header, and TotalSegmentator is applied to label 117 structures; slices with vertebra index $=29$ are then selected as L3 (Ahmed et al., 19 Mar 2025). This makes slice localization an explicit computational stage rather than an informal precondition.

MRI-based L3 workflows can still depend on expert localization. In DAFS Express, all In-Phase Dixon MRI scans are imported, metadata such as pixel spacing and slice thickness are read, and a trained rater visually identifies and extracts the true axial slice at the L3 vertebral level before automated segmentation proceeds (Akella et al., 2024). The MRI workflow also includes resampling to a consistent in-plane resolution and definition of a loose abdominal ROI to accelerate segmentation (Akella et al., 2024).

A separate CT sarcopenia workflow used manual L3 identification by a trained radiologist, who selected the single transverse slice passing through the inferior endplate of the third lumbar vertebra and exported it as a 2-D bitmap image (Bhardwaj et al., 24 Aug 2025). Preprocessing then re-oriented all 3-D volumes to RAS, resampled voxel spacing to $1 \text{ mm}\times1 \text{ mm}\times1 \text{ mm}$, clipped HU values to $[-175,+250]$, linearly mapped intensities to $[0,1]$, and extracted the L3 slice for downstream processing (Bhardwaj et al., 24 Aug 2025).

In longitudinal ultrasound, the localization problem is different. Skeletal muscle cross-sectional area measurement over time consists of three steps: retrieving the same anatomical slice at each visit, segmenting the muscle boundary, and converting that segmentation into an area measurement (Kerdegari et al., 2023). The first step is especially critical because manual ultrasound measurements are time-consuming and suffer large inter- and intra-operator variability due to inconsistent probe placement, angulation, and subjective tracing of the muscle border (Kerdegari et al., 2023).

3. Computational measurement pipelines

Ultrasound SMA tracking in ICU patients has been formulated as a retrieval-plus-segmentation problem. A self-supervised contrastive model inspired by SimCLR augments each frame $X$ twice by random crop and flip to form a positive pair $(X_i,X_j)$. A shared encoder $f(\cdot)$, consisting of four $r\approx0.95$0 convolution layers with ReLU and two max-pooling layers, maps each view to a representation $r\approx0.95$1, and a two-layer MLP projection head $r\approx0.95$2 with 512 units per layer yields $r\approx0.95$3. The contrastive loss is

$r\approx0.95$4

where $r\approx0.95$5 is cosine similarity and $r\approx0.95$6 is a temperature parameter. After 500 epochs of self-supervised pretraining, the encoder is frozen and its outputs feed a classifier with four dense layers of $r\approx0.95$7 units and softmax to predict whether two frames correspond to the same anatomical view (Kerdegari et al., 2023). The retrieved frame is then segmented by a standard U-Net trained on 1,000 manually annotated rectus-femoris cross-sections using Adam optimization, $r\approx0.95$8 regularization, and data augmentation (Kerdegari et al., 2023).

The CT oncology pipeline in SMAART-AI is fully automated from image retrieval through prognostic modeling. After DICOM retrieval and L3 identification, HU values are windowed to $r\approx0.95$9 to isolate muscle density, converted to PNG, and normalized to zero mean and unit variance. A 2D nnU-Net receives one $$\mathrm{SMI} \;=\; \frac{\mathrm{SMA}}{\bigl(\text{height in m}\bigr)^2} \quad \bigl[\mathrm{cm}^2/\mathrm{m}^2\bigr].$$0 normalized PNG slice and outputs a pixel-wise probability map and binary mask using a threshold of 0.5. The architecture is a 2D U-Net with 7 encoder and 7 decoder stages, skip connections, instance normalization, and LeakyReLU activations; two variants are described, PlainConvUNet and ResidualEncoderUNet, the latter using deeper residual blocks in the encoder. Training uses the nnU-Net default composite Dice plus cross-entropy loss and stochastic gradient descent with initial learning rate $$\mathrm{SMI} \;=\; \frac{\mathrm{SMA}}{\bigl(\text{height in m}\bigr)^2} \quad \bigl[\mathrm{cm}^2/\mathrm{m}^2\bigr].$$1, under 5-fold cross-validation, with average pixel probabilities across folds used at inference (Ahmed et al., 19 Mar 2025).

DAFS Express applies a related MRI segmentation strategy to single L3 slices. The software imports DICOM data, applies a proprietary intensity-normalization routine to standardize fat-water contrast, and uses a U-Net-style CNN to classify each pixel as SKM, VAT, or SAT. For SKM, the network output is thresholded to a binary mask, small isolated islands under 50 pixels are removed, and holes within the SKM mask are filled. The final mask is converted into a tag file, the cross-sectional area is computed, mean signal intensity within the SKM mask is recorded for potential fat-infiltration analyses, and a one-page PDF report is generated per slice (Akella et al., 2024).

A further CT workflow couples transfer learning for sarcopenia classification with self-supervised segmentation for quantitative SMA estimation. The classification arm replaces the final layer of Densenet-121, InceptionV3, InceptionResNetV2, and ResNet-50 backbones pretrained on RadImageNet with a two-node SoftMax head. The segmentation arm, termed SMIT, uses a Vision Transformer in a teacher-student self-distillation framework with masked image modeling, pretrained on more than 3,500 unlabeled CT scans and then fine-tuned on 79 labeled CT volumes with masks. This student network outputs a pixel-wise skeletal-muscle mask $$\mathrm{SMI} \;=\; \frac{\mathrm{SMA}}{\bigl(\text{height in m}\bigr)^2} \quad \bigl[\mathrm{cm}^2/\mathrm{m}^2\bigr].$$2 (Bhardwaj et al., 24 Aug 2025).

4. Quantification formulas, evaluation metrics, and reported performance

Across modalities, SMA is computed from a binary muscle mask and pixel geometry. In the ultrasound formulation,

$$\mathrm{SMI} \;=\; \frac{\mathrm{SMA}}{\bigl(\text{height in m}\bigr)^2} \quad \bigl[\mathrm{cm}^2/\mathrm{m}^2\bigr].$$3

where $$\mathrm{SMI} \;=\; \frac{\mathrm{SMA}}{\bigl(\text{height in m}\bigr)^2} \quad \bigl[\mathrm{cm}^2/\mathrm{m}^2\bigr].$$4 is the total number of pixels, $$\mathrm{SMI} \;=\; \frac{\mathrm{SMA}}{\bigl(\text{height in m}\bigr)^2} \quad \bigl[\mathrm{cm}^2/\mathrm{m}^2\bigr].$$5 indicates whether pixel $$\mathrm{SMI} \;=\; \frac{\mathrm{SMA}}{\bigl(\text{height in m}\bigr)^2} \quad \bigl[\mathrm{cm}^2/\mathrm{m}^2\bigr].$$6 belongs to muscle, and $$\mathrm{SMI} \;=\; \frac{\mathrm{SMA}}{\bigl(\text{height in m}\bigr)^2} \quad \bigl[\mathrm{cm}^2/\mathrm{m}^2\bigr].$$7 is the known pixel spacing; no further coordinate transform is needed beyond pixel-to-millimeter conversion (Kerdegari et al., 2023). The MRI formulation is equivalent:

$$\mathrm{SMI} \;=\; \frac{\mathrm{SMA}}{\bigl(\text{height in m}\bigr)^2} \quad \bigl[\mathrm{cm}^2/\mathrm{m}^2\bigr].$$8

with $$\mathrm{SMI} \;=\; \frac{\mathrm{SMA}}{\bigl(\text{height in m}\bigr)^2} \quad \bigl[\mathrm{cm}^2/\mathrm{m}^2\bigr].$$9 obtained from DICOM metadata (Akella et al., 2024).

Reported quantitative results span retrieval, segmentation overlap, area error, and reliability metrics:

Setting	Reported results	Study
ICU ultrasound longitudinal exams	AUC $=29$0; precision $=29$1; recall $=29$2; F1 $=29$3; mean relative absolute CSA error $=29$4	(Kerdegari et al., 2023)
CT cancer cachexia pipeline	Mean Dice $=29$5; median absolute error $=29$6; uncertainty correlations: variance $=29$7, entropy $=29$8, CV $=29$9	(Ahmed et al., 19 Mar 2025)
MRI L3 body-composition analysis	Mean SKM area automated $1 \text{ mm}\times1 \text{ mm}\times1 \text{ mm}$0; manual $1 \text{ mm}\times1 \text{ mm}\times1 \text{ mm}$1; bias $1 \text{ mm}\times1 \text{ mm}\times1 \text{ mm}$2; Bland-Altman limits $1 \text{ mm}\times1 \text{ mm}\times1 \text{ mm}$3; Jaccard $1 \text{ mm}\times1 \text{ mm}\times1 \text{ mm}$4; Dice $1 \text{ mm}\times1 \text{ mm}\times1 \text{ mm}$5; ICC $1 \text{ mm}\times1 \text{ mm}\times1 \text{ mm}$6; runtime $1 \text{ mm}\times1 \text{ mm}\times1 \text{ mm}$7 s per DICOM	(Akella et al., 2024)
CT sarcopenia segmentation	Mean Dice $1 \text{ mm}\times1 \text{ mm}\times1 \text{ mm}$8; mean SMA error $1 \text{ mm}\times1 \text{ mm}\times1 \text{ mm}$9 percentage-points; average fold errors from $[-175,+250]$0 to $[-175,+250]$1; after cut-off thresholding, held-out sarcopenia detection accuracy $[-175,+250]$2	(Bhardwaj et al., 24 Aug 2025)

The overlap metrics are standard. The Dice similarity coefficient is written as

$[-175,+250]$3

where $[-175,+250]$4 is the predicted mask and $[-175,+250]$5 is the manual mask (Ahmed et al., 19 Mar 2025). The MRI study additionally reports the Jaccard coefficient, Intraclass Correlation Coefficient under a two-way mixed absolute-agreement model, and Bland-Altman bias and limits of agreement, thereby distinguishing pixel overlap from agreement in area measurements (Akella et al., 2024). This distinction matters because reliable SMA quantification is not exhausted by boundary overlap alone.

5. Clinical roles in critical illness, sarcopenia, and cancer cachexia

In critical care, skeletal muscle atrophy is common in ICU patients who spend long periods in bed, and muscle mass must be recovered through physiotherapy before discharge. Ultrasound is frequently used to assess recovery by measuring muscle size over time, but manual measurements suffer from large variability. Automated retrieval of the same cross-section at each visit, combined with automated U-Net segmentation, reduces random and systematic errors in longitudinal SMA tracking and enables more precise monitoring of muscle atrophy and recovery (Kerdegari et al., 2023). The same study states that objective, reproducible quantification may guide the timing and intensity of physiotherapy, help predict functional outcomes, and facilitate large-scale studies of muscle wasting interventions (Kerdegari et al., 2023).

In oncology, progressive loss of SMA is presented as a hallmark of cancer cachexia, a multifactorial syndrome characterized by involuntary weight and muscle loss. Early detection of declining SMA can guide nutritional and exercise interventions, improve treatment tolerance, and is strongly prognostic of survival (Ahmed et al., 19 Mar 2025). SMAART-AI extends SMA estimation into a multimodal cachexia-prediction workflow by combining SMA, SMI, BMI, and clinical variables such as age, sex, race, ethnicity, and cancer stage in a multi-layer perceptron for binary cachexia classification at diagnosis. The described architecture uses three hidden linear layers with LeakyReLU and dropout followed by a sigmoid output; reported evaluation includes $[-175,+250]$6 precision for predicting cachexia, $[-175,+250]$7 overall accuracy, and an F1-score of $[-175,+250]$8 (Ahmed et al., 19 Mar 2025).

For sarcopenia detection in CT, one study reports that quantitative assessment of SMA is more precise and informative than qualitative assessment models for detecting sarcopenia (Bhardwaj et al., 24 Aug 2025). After segmentation, SMA can be thresholded by published cut-offs of male $[-175,+250]$9 and female $[0,1]$0, yielding $[0,1]$1 sarcopenia detection accuracy on held-out test sets in that dataset (Bhardwaj et al., 24 Aug 2025). This frames SMA not merely as a descriptive measurement but as a quantitative intermediate phenotype for diagnostic stratification.

6. Reliability, variability, and recurring methodological issues

The principal threats to reliable SMA estimation differ by modality. In ultrasound, variability arises from inconsistent probe placement, angulation, and subjective tracing of the muscle border; the retrieval stage is therefore indispensable, not optional (Kerdegari et al., 2023). In MRI, documented variability sources include poor image quality such as motion blur and stitching artifacts, unusual anatomy including hip prostheses and severe scoliosis, varying voxel sizes, and rater inconsistency during manual correction, although the reported ICC of 0.998 indicates that rater-induced disagreement was largely mitigated in that study (Akella et al., 2024).

Uncertainty quantification has become an explicit mechanism for reliability control in automated CT SMA pipelines. SMAART-AI computes variance, entropy, and coefficient of variation from an ensemble of models, and flags cases above dataset-specific uncertainty thresholds for expert correction. The reported Pearson correlations between these uncertainty measures and absolute SMA error—approximately $[0,1]$2 for variance, $[0,1]$3 for entropy, and $[0,1]$4 for coefficient of variation—indicate that elevated uncertainty reliably flags likely high-error cases (Ahmed et al., 19 Mar 2025). This directly addresses the misconception that high mean segmentation performance eliminates the need for human review.

Another recurring methodological issue is conflation of slice selection with segmentation accuracy. The workflows summarized here make slice or view localization a separate stage: TotalSegmentator-based L3 identification in CT, expert L3 selection in MRI and in one CT study, and contrastive retrieval of corresponding ultrasound views over time (Ahmed et al., 19 Mar 2025, Akella et al., 2024, Bhardwaj et al., 24 Aug 2025, Kerdegari et al., 2023). A plausible implication is that reproducible SMA estimation depends jointly on anatomical standardization, acquisition consistency, and segmentation quality, rather than on the segmentation network alone.