Discussion

This study is, to our knowledge, the first to systematically compare two different deep learning-based segmentation–radiomics paradigms (object-detection-based YOLOv8-hybrid and fully-convolutional nnU-Net v2) within the same methodological framework for forensic maxillary sinus analysis. The YOLOv8-hybrid pipeline demonstrated exceptional consistency (mAP@50 = 98.19%, CV = 0.77%) and high throughput. nnU-Net v2 provided high pixel precision (median DSC > 0.86) and statistically significant superior classification performance (AUC = 0.927 [0.881, 0.964] vs. 0.893 [0.841, 0.938], DeLong p = 0.024, Cohen’s d = 0.48). Both systems provided robust age estimation (MAE ≈ 7.2 years, 95% CI: [5.98, 8.49]).

Segmentation performance

YOLOv8’s CV = 0.77% showed robust generalization across different data subsets. In the literature, YOLO architectures are widely used in high-throughput scenarios [40]. nnU-Net v2 provided pixel-level precision with median DSC > 0.86, consistent with its state-of-the-art performance (mean DSC 0.89) in the Medical Segmentation Decathlon [22, 54]. However, the significant discrepancy between mean (~ 0.69) and median (~ 0.86) DSC indicates a right-skewed distribution with a long tail of failure cases (n = 54, 17.6%). These failures were systematically associated with aplastic/hypoplastic sinuses (38.9%), low image quality (33.3%), metal artifacts (27.8%), and positional variation (13.0%). Anatomical variants such as sinus septa [41] may further complicate automated segmentation. Future studies should explore artifact-robust architectures, foundation models such as SAM [53], or dedicated pre-filtering modules.

Sex classification and sexual dimorphism

nnU-Net-XGBoost achieved 85.27% accuracy and Cohen’s Kappa = 0.703 (“substantial agreement”), exceeding previous panoramic studies: Uthman et al. [11] 83.3%, Sharma et al. [28] 70%, Divyadharsini & Uma Maheswari [48] 72.5%, and de Queiroz et al. [49] 69.2%, and Gurses et al. [58] 76.8% (ML-based). Males showed significantly larger sinuses on both sides (p < 0.001). Cohen’s d = 0.52 (right sinus, medium effect) supports post-pubertal androgen-driven pneumatization [10]. The prominence of bilateral area difference in SHAP analysis (0.42) quantitatively confirms maxillary sinus asymmetry as a critical forensic biomarker.

Age estimation

MAE = 7.20 years (nnU-Net) and 7.30 years (YOLOv8) exceeded previous studies: Cameriere et al. [29] and Gulsahi et al. [30] (MAE ≈ 10 + years). Vila-Blanco et al. [42] demonstrated deep neural networks for chronological age estimation from full panoramic radiographs; however, our study is the first to target specifically sinus-derived features. Bland-Altman analysis confirmed no systematic bias (mean: −0.34 years). In forensic anthropology, ± 10 years is considered “acceptable” [3]; 91.8% of predictions fell within ± 15 years.

Radiomics contribution

The integration of > 120 morphometric and radiomic features significantly exceeded traditional 4–12 feature approaches. GLCM contrast reflects trabecular heterogeneity differences between sexes; GLRLM run percentage captures age-related deterioration in bone alignment.

Biological Signal Strength and Forensic Utility: We acknowledge that maxillary sinus morphology provides a moderate rather than strong biological signal for sex and age estimation, as noted in previous systematic reviews [60]. The effect sizes observed in this study (Cohen’s d = 0.52 for right sinus area, AUC = 0.927 for sex classification) are consistent with the known modest sexual dimorphism of paranasal sinuses [10, 12, 31]. However, the forensic utility of the maxillary sinus lies precisely in its taphonomic resilience — it often remains intact in severely decomposed, burned, or fragmented remains where stronger indicators such as the pelvis or pubic symphysis are unavailable [9, 46, 47]. Our pipelines are not proposed as standalone forensic identification tools but as supplementary components within multi-indicator biological profiling frameworks, consistent with contemporary forensic anthropological practice [1, 3]. The achieved AUC of 0.927 represents a statistically significant improvement over prior manual measurement-based methods (69–83%) [11, 28, 48, 49] and provides quantified posterior probabilities with defined error margins suitable for forensic reporting. Furthermore, the radiomics approach captures texture-level information (GLCM, GLRLM, GLSZM) that is inaccessible to traditional linear measurements, potentially explaining the performance gain over conventional approaches [14, 15, 59].

Inter-observer reliability

The addition of dual-annotator verification addresses a common limitation in medical AI studies. ICC values of 0.94–0.97 (“excellent”) and mean DSC of 0.934 between observers demonstrate that the ground truth annotations are highly reliable and reproducible. This strengthens the validity of all downstream model evaluations.

External validation

Temporal external validation on 50 independent images from a distinct patient cohort (2014–2016) confirmed the generalizability of both pipelines. The observed performance (nnU-Net sex accuracy: 82%, age MAE: 7.85 years) was consistent with primary test set results, with minor degradation expected from a smaller sample and inter-cohort variability. YOLOv8 showed higher segmentation success rates (92% vs. 86% with DSC > 0.5), consistent with its robustness advantage.

Clinical implications and architecture selection guide

Based on our findings, we propose the following architecture selection framework:

(1) High-throughput scenarios (mass disaster triage): YOLOv8-Hybrid (10× fewer parameters, CV = 0.77%).

(2) Maximum precision applications (forensic reporting, court evidence): nnU-Net v2 (AUC = 0.927, p = 0.024).

(3) Hardware-constrained environments (mobile, field deployment): YOLOv8-Hybrid (3.2 M vs. 31.2 M parameters).

Model interpretability

SHAP analysis provides explicit reporting of which features contribute to predictions, positioning these models not only as classifiers but as forensic opinion generators with defined error margins. This is consistent with the requirement for explainable known error rates in forensic evidence.

Limitations

(a) Single center/population: multi-ethnic, multi-center validation is required.(b) Retrospective design: prospective clinical studies are needed.(c) 2D modality: CBCT provides richer 3D information, but panoramic is more accessible and lower dose.(d) Adult age range (18–81): including pediatric groups may improve accuracy [43].(e) Cross-sectional design: longitudinal data can directly show aging changes.(f) Trauma and pathology not included: forensic cases may require trauma detection modules.(g) External validation sample size: larger external cohorts from multiple centers would strengthen generalizability claims.