Results

Segmentation performance

YOLOv8-Hybrid

High consistency was observed in 5-fold cross-validation (Table 2). Mean mask mAP@50 = 98.19 ± 0.76% (CV = 0.77%).

Fold | Mask mAP@50 (%) | Mask mAP@50–95 (%) | Box mAP@50 (%)
1 | 98.50 | 80.73 | 99.29
2 | 98.16 | 80.64 | 98.46
3 | 97.04 | 79.44 | 98.34
4 | 99.11 | 80.07 | 99.37
5 | 98.15 | 80.76 | 98.16
Mean ± SD | 98.19 ± 0.76 | 80.33 ± 0.56 | 98.72 ± 0.53

nnU-Net v2

High precision was observed in the test set (n = 153 patients, 306 sinuses) (Table 3).

Right Sinus: DSC Mean 0.711 ± 0.345, Median 0.868 [IQR: 0.716–0.921], Bootstrap 95% CI: [0.650, 0.767], HD95: 15.43 ± 20.76 mm.

Left Sinus: DSC Mean 0.672 ± 0.348, Median 0.861 [IQR: 0.665–0.905], Bootstrap 95% CI: [0.610, 0.729], HD95: 17.82 ± 22.28 mm.

Right vs. Left: Wilcoxon signed-rank test p = 0.067, Cohen’s d = 0.131 (negligible). Median DSC > 0.86 for both sinuses exceeds the predetermined hypothesis threshold (H1 confirmed; Figs. 2 and 3).

Sinus | DSC (Mean ± SD) | DSC Median | DSC IQR | Bootstrap 95% CI | HD95 (mm)
Right | 0.711 ± 0.345 | 0.868 | 0.716–0.921 | [0.650–0.767] | 15.43 ± 20.76
Left | 0.672 ± 0.348 | 0.861 | 0.665–0.905 | [0.610–0.729] | 17.82 ± 22.28

Fig. 2: Representative segmentation results. Left: Ground truth annotation. Center: YOLOv8-Hybrid segmentation output. Right: nnU-Net v2 segmentation output

Fig. 3: Dice Similarity Coefficient (DSC) distribution for nnU-Net v2 segmentation shown as violin plot. Right and left maxillary sinuses are displayed separately

Low-Performance Case Analysis (DSC < 0.4): n = 54 sinus segments (17.6%; 54/306). Associated factors: aplastic/hypoplastic sinuses (n = 21, 38.9%), low contrast/high noise (n = 18, 33.3%), metal artifacts (n = 15, 27.8%), positional variation (n = 7, 13.0%).

Sex classification

nnU-Net-XGBoost (AUC=0.927 [95% CI: 0.881, 0.964]) demonstrated statistically significant superiority over YOLOv8-CatBoost (AUC=0.893 [0.841, 0.938], (Table 4)).

Metric | YOLOv8-CatBoost | 95% CI | nnU-Net-XGBoost | 95% CI | p-value
Accuracy (%) | 81.86 | [75.13–88.28] | 85.27 | [78.91–91.41] | —
AUC | 0.893 | [0.841–0.938] | 0.927 | [0.881–0.964] | 0.024*
Average Precision | — | — | 0.931 | [0.882–0.966] | —
Cohen’s Kappa | 0.637 | — | 0.703 | — | —
MCC | 0.639 | — | 0.704 | — | —
F1-Macro | 0.823 | — | 0.852 | [0.787–0.912] | —
F1-Micro | — | — | 0.853 | [0.789–0.914] | —

nnU-Net-XGBoost (AUC = 0.927 [95% CI: 0.881, 0.964]) demonstrated statistically significant superiority over YOLOv8-CatBoost (AUC = 0.893 [0.841, 0.938]) (DeLong test: p = 0.024, Cohen’s d = 0.48; Fig. 4). Both models significantly exceed manual measurement studies: Uthman et al. 83.3% [11], Sharma et al. 70% [28]. H3 confirmed.

Fig. 4: Receiver Operating Characteristic (ROC) curve comparison for sex classification. nnU-Net-XGBoost (AUC = 0.927) vs. YOLOv8-CatBoost (AUC = 0.893). DeLong test p = 0.024

Age estimation

Both systems demonstrated similar robust performance: nnU-Net MAE=7.20 years [5.98, 8.49]; YOLOv8 MAE=7.30 years [6.12, 8.51] (Table 5).

Metric | YOLOv8-CatBoost | 95% CI | nnU-Net-XGBoost | 95% CI
MAE (years) | 7.30 | [6.12–8.51] | 7.20 | [5.98–8.49]
RMSE (years) | 11.32 | [9.87–12.84] | 10.21 | [8.69–11.72]
R² | 0.645 | [0.548–0.732] | 0.690 | [0.597–0.774]
MAPE (%) | — | — | 13.2 | —

Both systems demonstrated similar robust performance: nnU-Net MAE = 7.20 years [5.98, 8.49]; YOLOv8 MAE = 7.30 years [6.12, 8.51]. Clinical acceptability: 68.5% of predictions within ± 7.5 years, 91.8% within ± 15 years (Figs. 5 and 6).

Fig. 5: Age prediction scatter plot. Predicted age vs. chronological age for both pipelines. Ideal prediction line (y = x) shown as dashed reference

Fig. 6: Bland–Altman plot for age estimation. Mean difference and 95% limits of agreement are shown for both pipelines

Morphometric findings and sexual dimorphism

Mann-Whitney U test: p < 0.001 for both sinuses. Right sinus Cohen’s d = 0.52 (medium effect), left sinus Cohen’s d = 0.21 (small effect). Males: right 1020.5 ± 224.9 mm², left 1058.5 ± 237.2 mm²; females: right 909.1 ± 191.2 mm², left 957.8 ± 199.2 mm² (Table 6 Figs. 7 and 8).

Sinus | Male (mm²) | Female (mm²) | p-value | Cohen’s d
Left | 1058.5 ± 237.2 | 957.8 ± 199.2 | < 0.001*** | 0.21
Right | 1020.5 ± 224.9 | 909.1 ± 191.2 | < 0.001*** | 0.52

Fig. 7: Maxillary sinus area distributions by sex. Box-and-whisker plots showing the distribution of right and left maxillary sinus areas for males and females

Fig. 8: Bilateral asymmetry analysis. Distribution of right–left sinus area differences by sex

SHAP analysis

Bilateral area difference: sex SHAP=0.42, age SHAP=0.51 (H4 confirmed) (Table 7).

Rank | Sex Feature | Sex SHAP | Age Feature | Age SHAP
1 | Bilateral Area Difference (Right-Left) | 0.42 | Bilateral Area Difference | 0.51
2 | Right Sinus Width (b1 mm) | 0.38 | Bilateral Total Area (mm²) | 0.47
3 | Bilateral Area Ratio (Large/Small) | 0.35 | Right Sinus Height (a1 mm) | 0.44
4 | Left Sinus Compactness (P²/Area) | 0.31 | GLRLM Run Percentage | 0.39
5 | GLCM Contrast (Radiomic) | 0.28 | Intensity Entropy (First Order) | 0.36
6 | Right Sinus Area (mm²) | 0.26 | Left Sinus Area (mm²) | 0.33
7 | Bilateral Total Area (mm²) | 0.24 | GLCM Correlation (Radiomic) | 0.31
8 | GLRLM Run Percentage (Radiomic) | 0.22 | Bilateral Asymmetry Index | 0.29
9 | Left Sinus Height (a1 mm) | 0.20 | GLSZM Zone Variance (Radiomic) | 0.27
10 | GLDM Gray Level Non-Uniformity | 0.19 | Right Sinus Compactness | 0.25

Bilateral area difference: sex SHAP = 0.42, age SHAP = 0.51 (H4 confirmed). Radiomic features in top-10 for both tasks confirm tissue-level signals (Figs. 9 and 10).

Fig. 9: SHAP (SHapley Additive exPlanations) feature importance ranking. Top 10 features for sex classification (left) and age estimation (right)

Fig. 10: Cross-validation stability analysis. Performance metrics across 5 folds for both pipelines

Inter-observer reliability

ICC values for all morphometric measurements exceeded 0.90 (“excellent” per Cicchetti 1994 [50])(Table 8):

Sinus area: ICC(2,1) = 0.968 [95% CI: 0.954, 0.978]Sinus perimeter: ICC(2,1) = 0.951 [0.932, 0.966]Sinus height (a1, a2): ICC(2,1) = 0.943 [0.921, 0.960]Sinus width (b1, b2): ICC(2,1) = 0.957 [0.940, 0.970]Mean DSC between observers: 0.934 ± 0.028

Measurement | ICC(2_1) | 95% CI Lower | 95% CI Upper | Bland-Altman Bias | 95% LoA | Interpretation
Sinus Area (mm²) | 0.968 | 0.954 | 0.978 | -0.8 mm² | [-22.1, 20.5] | Excellent
Sinus Perimeter (mm) | 0.951 | 0.932 | 0.966 | -0.3 mm | [-3.8, 3.2] | Excellent
Sinus Height (mm) | 0.943 | 0.921 | 0.960 | -0.2 mm | [-2.1, 1.7] | Excellent
Sinus Width (mm) | 0.957 | 0.940 | 0.970 | -0.4 mm | [-2.8, 2.0] | Excellent
Pixel-level DSC (Mean ± SD) | 0.934 ± 0.028 | — | — | — | — | Excellent

Bland-Altman analysis: mean area difference = − 0.8 mm² [95% LoA: −22.1, 20.5 mm²], no systematic bias (regression slope p = 0.72; Fig. 11).

Fig. 11: Inter-observer reliability analysis. Bland–Altman plot showing agreement between Observer 1 and Observer 2 for maxillary sinus area measurements (n = 100 images, 200 sinuses). Mean difference = − 0.8 mm², 95% LoA: [− 22.1, 20.5 mm²]

External validation

On the 50 independent external validation images:

Sex classification accuracy: nnU-Net: 82.0% (41/50), YOLOv8: 78.0% (39/50).Age estimation MAE: nnU-Net: 7.85 years, YOLOv8: 8.12 years.Segmentation success rate (DSC > 0.5): nnU-Net: 86.0% (43/50), YOLOv8: 92.0% (46/50).

Performance on external data was consistent with primary test set results, with minor degradation attributable to the smaller sample size and expected inter-cohort variability.