**Background:** Malnutrition is highly prevalent in cancer patients and is associated with reduced treatment sensitivity, quality of life, and increased mortality. Current screening tools vary across regions and rely on subjective assessments. Objective measures like DXA, CT, and MRI are accurate but costly and not widely used. 3D facial imaging has been applied in acromegaly, genetic syndromes, and skin lesions, but its utility in malnutrition assessment had not been explored. This study aimed to evaluate whether 3D facial image recognition could identify depression indices in the temporal and periorbital regions that correlate with malnutrition phenotypes in cancer patients.
**Methods:** This cross-sectional pilot study recruited 53 hospitalized patients with advanced digestive system malignancies receiving maintenance chemotherapy at Peking Union Medical College Hospital. Inclusion criteria: Han race, age > 18 years, NRS 2002 score > 3, voluntary participation. Exclusion criteria included facial plastic surgery/trauma, head/neck radiotherapy, diseases causing facial changes (acromegaly, hyperthyroidism), high-dose glucocorticoid use, and facial/limb edema. Within 24 hours of admission, two trained physicians performed NRS 2002 and subjective global assessment (SGA) nutritional assessments, physical examination (height, weight, upper arm circumference [AC], calf circumference [CC], handgrip strength [HGS]), and body composition analysis via bioelectrical impedance analysis (InBody S10). Fat-free mass index (FFMI) and fat mass index (FMI) were calculated. 3D facial images were captured using the Antera 3D® system, which uses LEDs of different wavelengths and computer-assisted 3D skin surface reconstruction. Depression indices (volume, affected area, maximum depth) were quantified for the temporal region (diameter 30.8 mm) and periorbital region (diameter 54.1 mm) using Antera Pro software v2.8.2. Malnutrition phenotype groups were defined using established cutoffs: low AC (<27 cm men, <25 cm women), low CC (<33 cm men, <32 cm women), low HGS (<28 kg men, <18 kg women), low FFMI (<17 kg/m² men, <15 kg/m² women), and low FMI (<7.7 kg/m² men, <5 kg/m² women). Statistical analyses included Spearman correlations, one-way ANOVA, Kruskal-Wallis H-test, independent samples t-test, and Mann-Whitney U-test, with significance set at p < 0.05.
**Key Results:** Among 53 patients, SGA classified 30 as well-nourished (SGA A), 15 as mild-to-moderate malnutrition (SGA B), and 8 as severe malnutrition (SGA C). Age, sex, and tumor location were evenly distributed. Most tumors were gastrointestinal (74%). BMI, AC, CC, fat mass, FMI, and percent body fat (PBF) showed significant differences across SGA groups with a downward trend as malnutrition worsened (BMI: 24.12 vs 23.07 vs 18.84 kg/m², p < 0.001; AC: 28.50 vs 29.00 vs 24.50 cm, p = 0.006; CC: 36.95 vs 35.58 vs 32.73 cm, p = 0.003; FAT: 16.05 vs 13.35 vs 6.99 kg, p = 0.004; FMI: 5.68 vs 4.59 vs 2.47 kg/m, p = 0.003; PBF: 23.03 vs 19.29 vs 13.09%, p = 0.009). HGS and FFMI did not differ significantly. Temporal depression volume showed a grading trend across SGA groups (71.05 vs 75.59 vs 99.66 mm³, p = 0.068), and temporal depression area was significantly different (244.71 vs 259.21 vs 315.10 mm², p = 0.049). Temporal depression volume was significantly negatively correlated with AC (r = -0.293, p = 0.033) and CC (r = -0.285, p = 0.038). Periorbital depression volume and area were significantly negatively correlated with FMI (r = -0.273, p = 0.048; r = -0.304, p = 0.026) and PBF (r = -0.317, p = 0.021; r = -0.364, p = 0.007). Patients with muscle loss phenotypes had significantly higher temporal depression volume and area (low AC: 93.92 vs 70.45 mm³, p = 0.015; low CC: 100.26 vs 73.06 mm³, p = 0.031; low FFMI: 103.13 vs 70.49 mm³, p = 0.002). Patients with low FMI had significantly higher periorbital depression volume (761.37 vs 648.69 mm³, p = 0.003) and area (1106.34 vs 1006.60 mm², p = 0.001).
**Clinical Implications:** This study provides proof-of-concept that 3D facial image recognition can non-invasively capture depression indices in the temporal and periorbital regions that correlate with objective measures of muscle and fat loss in malnourished cancer patients. Temporal depressions reflect muscle mass loss, while periorbital depressions reflect subcutaneous fat loss, consistent with nutrition-focused physical assessment guidelines. The technology is simple, accessible (potentially via mobile phone applications), and could enable early malnutrition screening in settings where DXA, CT, or MRI are unavailable, such as community hospitals and nursing facilities. Limitations include the small sample size (n=53), use of BIA instead of gold-standard DXA/CT/MRI, and relatively modest correlation coefficients (0.2–0.4), suggesting that future studies with larger datasets and advanced machine learning techniques are needed to improve accuracy and validate these findings.