Study design and patient cohorts

This retrospective cohort study was conducted in two parts, utilizing two distinct patient cohorts from the Department of Oral and Maxillofacial Surgery at Osaka University Dental Hospital. The study protocol was approved by the Institutional Review Board of Osaka University Dental Hospital (Approval No. H29-E19) and was conducted in accordance with the Declaration of Helsinki. The requirement for written informed consent was waived by the ethics committee due to the retrospective nature of the analysis.

First, a model development cohort of 348 patients who underwent surgery under general anesthesia between January 2017 and December 2020 was used to train and validate the deep learning model. These patients had undergone preoperative head and neck CT imaging and bioelectrical impedance analysis (BIA). Exclusion criteria were age < 20 years, presence of infections that could affect nutritional status, and a history of head and neck dysplasia. This cohort was used to establish the ground truth for muscle segmentation and to determine the optimal cutoff values for sarcopenia based on muscle area. The characteristics of the model development cohort are summarized in Table 1.

Table 1 The data for setting the cutoff values

Second, a separate prognostic validation cohort of 247 patients with primary oral cancer who received treatment between January 2009 and December 2020 was used to evaluate the clinical and prognostic significance of the automated measurements. All patients had preoperative head and neck CT scans. Exclusion criteria for this cohort were direct tumor invasion into the masseter muscle, prior treatment for oral cancer, and age < 20 years, to ensure that measurements were not confounded by tumor presence or developmental changes. The characteristics of the prognostic validation cohort are summarized in Table 2.

CT image acquisition and ground truth segmentation

Pre-treatment head and neck CT scans were acquired using a multi-detector CT scanner. Imaging parameters were standardized as follows: 120 kVp, 200–330 mA, and slice thickness of 2.5–5.0 mm. Both contrast-enhanced and non-contrast scans were included.

For all patients in the model development cohort (n = 348), the ground truth segmentation of the masseter muscle was performed manually. The cross-sectional area of both the left and right masseter muscles was delineated at the axial slice corresponding to the floor of the maxillary sinus. This anatomical level was selected as a readily identifiable landmark to facilitate reproducible slice selection; additionally, when feasible, this level may reduce the influence of dental metal artifacts compared with more caudal mandibular levels. CT-based cross-sectional area of the jaw muscles has been shown to correlate with their physiological cross-sectional area [23]. The segmentation was performed by a trained oral and maxillofacial surgeon using Synapse VINCENT® v5.3 (FUJIFILM, Tokyo, Japan). To ensure accuracy and minimize bias, all manual segmentations were reviewed and confirmed by a board-certified oral and maxillofacial radiologist with over 10 years of experience.

Deep learning model for automated segmentationU-Net architecture

We developed our automated segmentation model based on the U-Net architecture, a CNN renowned for its efficacy in biomedical image segmentation [17]. The U-Net consists of a contracting path (encoder) to capture contextual information and a symmetric expanding path (decoder) to enable precise localization. Our implementation was based on the nnU-Net framework, a robust, self-configuring version of U-Net that adapts its architecture and hyperparameters to the specific dataset, ensuring optimal performance [18]. The model architecture included 19 convolutional layers with Rectified Linear Unit (ReLU) activation, four max-pooling layers for downsampling, and corresponding up-sampling layers in the decoder path. The input images were resized to 256 × 256 pixels for model training.

Model training and validation

The model development cohort (n = 348) was randomly partitioned into training (80%, n = 278), validation (10%, n = 35), and internal testing (10%, n = 35) sets. To enhance the model’s robustness and prevent overfitting, we employed an extensive data augmentation strategy, including random rotations (± 15 degrees), scaling (± 10%), elastic deformations, and horizontal flipping.

The model was trained for 100 epochs with a batch size of 32. We used the Adam optimizer with an initial learning rate of 0.001, which was adaptively reduced on a plateau of the validation loss. A combination of Dice loss and binary cross-entropy was used as the loss function to handle the class imbalance and optimize for segmentation overlap. Training was performed on an NVIDIA Tesla V100 GPU with 32 GB of memory, using the PyTorch framework.

Measurement of muscle area and nutritional indices

The total masseter muscle cross-sectional area (MMA) was calculated as the sum of the areas of the left and right masseter muscles. For the automated method, the area was calculated from the segmentation masks generated by our trained U-Net model (termed AI-MMA) using ImageJ software (v1.53, National Institutes of Health, USA).

In the model development cohort, the skeletal muscle index (SMI) was measured using an InBody570™ BIA device (InBody Japan, Tokyo, Japan) to serve as a reference for systemic muscle mass. The MMA cutoff values for sarcopenia were determined using receiver operating characteristic (ROC) curve analysis, with the established SMI diagnostic criteria from the Asian Working Group for Sarcopenia (AWGS 2019) as the reference standard (< 7.0 kg/m2 for males, < 5.7 kg/m2 for females) [24].

In the prognostic validation cohort, we collected data on various nutrition-related factors, including C-reactive protein (CRP), albumin (Alb), body mass index (BMI), and several prognostic scores such as the CRP/Alb ratio (CAR), neutrophil/lymphocyte ratio (NLR), and the Controlling Nutritional Status (CONUT) score [25, 26].

Statistical analysis

All statistical analyses were performed using EZR (Saitama Medical Center, Jichi Medical University, Japan), a graphical user interface for R (The R Foundation for Statistical Computing). A p-value < 0.05 was considered statistically significant.

The segmentation performance of the deep learning model was evaluated on the internal test set by comparing the AI-MMA masks to the manual ground truth masks using the Dice Similarity Coefficient (DSC). The correlation between AI-MMA and manually measured MMA was assessed using Pearson’s correlation coefficient and visualized with a Bland–Altman plot.

Overall survival (OS), defined as the time from primary treatment to death from any cause, was the primary endpoint. The prognostic significance of low AI-MMA was assessed in the validation cohort. Kaplan–Meier curves were generated to visualize survival differences, and the log-rank test was used for comparison. Univariate and multivariate Cox proportional hazards regression models were used to identify independent prognostic factors for OS. Variables with a p-value < 0.10 in the univariate analysis, along with clinically relevant factors such as age, sex, and tumor stage (UICC 8th edition), were included in the multivariate model.

Ethical statements

The study was conducted in accordance with the Declaration of Helsinki and approved by the Institutional Review Board of Osaka University Dental Hospital (protocol code H29-E19). Patient consent was waived for this retrospective analysis. All patient data were fully anonymized prior to analysis to protect confidentiality.

Data availability

The datasets generated and/or analyzed during the current study are not publicly available due to privacy and ethical restrictions concerning patient confidentiality but are available from the corresponding author on reasonable request.