Radiomics meets sarcopenia: Machine learning-based multimodal modeling for esophageal cancer outcomes

Table 1 Six studies published between 2022 and 2024 combined sarcopenia and radiomics assessments to predict clinical outcomes in esophageal cancer

Ref.	Tumor type	Prognostic targets	Imaging time/IV contrast	Muscle/fat image level	Radiomics image (ROI)
Zhou et al[16]	Esophageal squamous cell carcinoma	PFS and OS	Pretreatment/no	L3 CT	Tumor ROI on PET and CT
Hinzpeter et al[17]	Esophageal or gastroesophageal cancer	Metastatic disease and OS	Pretreatment/no	L3 CT	Tumor ROI on PET and CT
Vogele et al[18]	Esophageal or gastric cancer	Sarcopenia and PD	Pretreatment and follow-up/yes	L3 CT	CT muscle ROIs on psoas major, quadratus lumborum, erector spinae
Iwashita et al[19]	Esophageal cancer	OS	Pretreatment/no	L3 CT	CT muscle ROIs on psoas, erector spinae, quadratus, lumborum, and abdominal wall muscles
Hinzpeter et al[20]	Metastatic esophageal and gastroesophageal cancer	PFS and OS	Pretreatment/no	L3 CT	Tumor ROI on PET and CT
Anconina et al[21]	Esophagogastric adenocarcinoma	RFS and OS	Pretreatment/no	L3 CT	Tumor ROI on PET and CT

CT: Computed tomography; IV: Intravenous; L3: The third lumbar vertebral level; OS: Overall survival; PD: Progressive disease; PET: Positron emission tomography; PFS: Progression-free survival; RFS: Relapse-free survival; ROI: Region of interest.

Table 2 Summary of the predictive models and collected features, including clinical data, body composition indices, and radiomics, used in the six reviewed studies

Ref.	Model	Clinical data	Body composition index	Radiomics features or index
Zhou et al[16]	Cox regression nomogram	Tumor stage, LVI, BMI, age, sex, number of PET-positive lymph nodes, treatment	SMI, VATI, SATI, and sarcopenia (SMI < 34.4 cm2/m2 in females, < 45.4 cm2/m2 in males)	Extracted 92 radiomic features, used LASSO for feature selection, calculated Rad-scores
Hinzpeter et al[17]	Metastatic disease: LGBM and RF, OS: Cox regression	Age, sex, race, BMI, ECOG score, histology (subtype and grade), disease stage (M0 vs M1), tumor stage, SUV/SUL parameters	Sarcopenia (SMI < 34.4 cm2/m2 in females, < 45.4 cm2/m2 in males)	Extracted 307 radiomic features, filtered features based on significant difference (P < 0.01) between groups
Vogele et al[18]	DT, KNN, RF	Sex, age, tumor type (gastric/esophageal), tumor stage, nodes stage, M stage, PD	TPA and TPI sarcopenia (the lowest gender-specific quartile of TPI values at each time point)	Extracted 85 radiomic features, used LASSO for feature selection
Iwashita et al[19]	Cox regression, nomogram	Age, sex, performance status, tumor location, chemotherapy, radiation dose, ENI, tumor, node, and metastasis stage, and various blood parameters (CRP, Albumin, NLR, PLR, PNI, mGPS)	SMI, PMI	Extracted 837 radiomic features, used LASSO-COX regression and VIF for feature selection
Hinzpeter et al[20]	Cox regression, clinical, clinical + SMI, clinical + SMI + CT, clinical + SMI + CT + PET	Age, sex, race, BMI, ECOG score, tumor stage, tumor histology, tumor grade, distance LN, pleura, metastasis organ, SUV/SUL parameters	Sarcopenia (SMI < 34.4 cm2/m2 in females, < 45.4 cm2/m2 in males)	No specificity description
Anconina et al[21]	Cox regression nomogram	Age, sex, race, BMI, ECOG score, tumor stage, nodes stage, LVI	Sarcopenia (SMI < 34.4 cm2/m2 in females, < 45.4 cm2/m2 in males)	Extracted 42 radiomic features, filtered features based on significant difference (P < 0.01) between groups

BMI: Body mass index; CRP: C-reactive protein; CT: Computed tomography; DT: Decision tree; ECOG: Eastern Cooperative Oncology Group; ENI: Elective lymph node irradiation; KNN: K-nearest neighbors; LASSO: Least absolute shrinkage and selection operator; LGBM: Light gradient boosting machine; LN: Lymph node; LVI: Lymphovascular invasion; M: Metastasis; mGPS: Modified Glasgow prognostic score; NLR: Neutrophil-lymphocyte ratio; OS: Overall survival; PD: Progressive disease; PET: Positron emission tomography; PLR: Platelet-lymphocyte ratio; PMI: Psoas muscle index; PNI: Prognostic nutritional index; RF: Random forest; SATI: Subcutaneous adipose tissue index; SMI: Skeletal muscle index; SUL: Standardized uptake value normalized by lean body mass; SUV: Standardized uptake value; TPA: Total psoas area; TPI: Total psoas index; VATI: Visceral adipose tissue index; VIF: Variance inflation factor.

Table 3 Summary of selected features, final or best-performing models, and predictive performance in the six reviewed studies

Ref.	Sample size	Selected features	Best/final model	Performance
Zhou et al[16]	91	Clinical: Stage, LVI, BMI. Body composition: SMI, VATI. Rad-score: PFS: Conventional_SUVbwmin_CT, shape_sphericity, GLCM_correlation_CT, GLRLM_LGRE_CT, GLCM_homogeneity_PET, GLRLM_LRE_PET, and GLZLM_SZHGE_PET; OS: Shape_volume(mL)_CT, shape_compacity_CT, GLCM_contrast_CT, GLZLM_SZECT, conventional_SUVbwmean_PET, shape_compacity_PET, GLRLM_SRHGE_PET, GLZLM_SZHGE_PET, and GLZLM_GLNU_PET	Cox regression and nomogram (clinical + body composition + Rad-score)	PFS: C-index/AUC = 0.810 (95%CI: 0.737-0.884). OS: C-index/AUC = 0.806 (95%CI: 0.720-0.891)
Hinzpeter et al[17]	243	Metastatic disease: Radiomics features, clinical data (age, sex, race, BMI, ECOG, histopathology, grade), SMI; OS: Clinical (ECOG ≥ 2, age ≥ 70, M1 disease), radiomics (CT: Shape volume, PET: GLZLM ZLNU)	Metastatic disease: LGBM. OS: Cox regression	Metastatic disease: Accuracy 80%, AUC 88%, recall 84%, precision 80%, F1-score 82%. OS: Kaplan-Meier survival analysis and hazard ratios showed statistical significance in the selected features; specific AUC/C-index not reported
Vogele et al[18]	83	Radiomics features (area, HMXG, HP10, HRMAD, IIQR, IMN, Intensity P10th, GCS)	RF	Sarcopenia: Accuracy 0.90 ± 0.03; AUC 0.96 ± 0.02. PD: Accuracy 0.88 ± 0.04; AUC 0.93 ± 0.04. Sarcopenia + PD: Accuracy 0.93 ± 0.04; AUC 0.97 ± 0.04
Iwashita et al[19]	98	Clinical (chemotherapy, prophylactic lymph node area irradiation, albumin, PNI, PLR, CRP/Alb ratio). Radiomics (L3wavelet.LHLfirstorderSkewness, L3wavelet.HLLgldmSmallDependenceLowGrayLevelEmphasis, L3wavelet.HLHfirstorderMean, L3wavelet.HLHglcmMCC, Lumbowavelet.HLLngtdmBusyness, Lumbowavelet.HLLngtdmContrast, AbsoriginalfirstorderUniformity, Abswavelet.HHHglcmImc2)	Cox regression and nomogram (clinical + radiomic)	OS: Validation dataset: Accuracy 75%, specificity 92%, sensitivity 75%, AUC = 0.86 (0.70-1.00), C-index = 0.88 (0.85-0.91)
Hinzpeter et al[20]	128	PFS: Clinical (ECOG, bone metastases), SMI, CT radiomics (NGLDM contrast), PET radiomics (shape volume, 70_Kurtosis). OS: Clinical (ECOG, bone metastases), SMI, CT radiomics (NGLDM Coarseness), PET radiomics (GLZLM_SZLGE)	Cox regression. OS: 6-36 months, clinical + SMI + CT + PET; at 3 month, clinical + SMI + CT; PFS: 3-21 months, clinical + SMI + CT + PET; 24-36 months, clinical + SMI	PFS: At 30 months, model (clinical + SMI) AUC = 0.86. OS: At 24 months, model (clinical + SMI + CT + PET) AUC = 0.88
Anconina et al[21]	145	RFS: Clinical (ECOG, tumor stage, nodes stage), sarcopenia, CT mean HU, PET kurtosis. OS: Clinical (ECOG, tumor stage, nodes stage), sarcopenia, CT mean HU	Cox regression and nomogram (clinical + sarcopenia + CT + PET)	RFS: At 24 months and 30 months, model (clinical + SMI + CT + PET) AUC = 0.81. OS: At 18 months and 24 months, model (clinical + SMI + CT) AUC = 0.80

AUC: Area under the curve; BMI: Body mass index; C-index: Concordance index; CI: Confidence interval; CRP/Alb: C-reactive protein/albumin ratio; CT: Computed tomography; ECOG: Eastern Cooperative Oncology Groups; GCS: Clustershade; GLCM: Gray-level co-occurrence matrix; GLNU: Gray level non-uniformity; GLRLM: Gray-level run-length matrix; GLZLM: Gray-level zone length matrix; HMXG: Maximum histogram gradient; HP10: Histogram P10^th; HRMAD: Histogram robust mean absolute deviation; HU: Hounsfield unit; IIQR: Intensity interquartile range; IMN: Intensity mean; LGBM: Light gradient boosting machine; LVI: Lymphovascular invasion; NGLDM: Neighborhood gray-level difference matrix; OS: Overall survival; PD: Progressive disease; PET: Positron emission tomography; PFS: Progression-free survival; PNI: Prognostic nutritional index; PLR: Platelet-lymphocyte ratio; RF: Random forest; RFS: Relapse-free survival; SMI: Skeletal muscle index; SUV: Standardized uptake value; SZHGE: Short-zone high gray-level emphasis; SZLGE: Short-zone low gray-level emphasis; VATI: Visceral adipose tissue index; ZLNU: Zone length non-uniformity.