Introduction

Forensic anthropology aims to determine the biological profile (sex, age, stature, ancestry) of unidentified human remains [1]. In the forensic context, the term “sex” refers to biological sex inferred from skeletal morphology and should not be confused with gender identity. Sex and age estimation play critical roles in mass disasters, war crimes, and forensic investigations by narrowing the search universe [2, 3]. Classical anthropometric approaches provide sex determination with high accuracy from the pelvis (90% accuracy) and cranial morphology (80–85% accuracy) [4]. Dental development [5], epiphyseal fusion, and pubic symphysis changes are considered the gold standard for age estimation [6]. However, in mass disasters, trauma, and advanced decomposition, these structures may be damaged [7, 8]. Analysis of protected anatomical regions through objective, reproducible, and automated methods is therefore necessary.

Maxillary sinuses carry forensic biomarker potential as durable structures resistant to postmortem changes in the skull [9, 46, 47]. Larger sinus volumes after puberty due to androgenic effects in males have been reported [10–12]. Biological sexual dimorphism in size and shape has been demonstrated in numerous studies. Panoramic radiography is used for both clinical and forensic purposes with low radiation dose (0.002–0.010 mSv; ~1% of CT) and widespread accessibility [13]. Radiomic textural metrics (gray-level co-occurrence matrix, gray-level run-length matrix) enhance age and sex signals by quantitatively capturing tissue heterogeneity [14, 15].

Deep learning has created a paradigm-shifting impact in medical image analysis [16]. Convolutional neural networks, including deep residual architectures [27], automatically learn complex patterns [17, 57]. Their application in forensic science enables automated biological profiling from skeletal remains [18]. Two different paradigms are prominent in deep learning methodology: the YOLO family provides high speed with single-stage detection and segmentation [19–21]; nnU-Net is accepted as the gold standard in medical segmentation [22, 23].

Previous studies on age estimation and sex classification from the maxillary sinus have mainly relied on manual linear/volumetric measurements and discriminant function analysis [10, 11, 28, 55, 60]. In contrast, recent years have reported high accuracy from machine learning/deep learning-based approaches for automated age and sex estimation from panoramic radiographs [29, 31, 43].

The unique aspect of this study is to develop and compare two deep learning-based segmentation–radiomics pipelines within the same methodological framework for age estimation and sex classification using maxillary sinus morphology from panoramic radiographs:

1. YOLOv8-hybrid: YOLOv8n-seg segmentation + U-Net boundary refinement + > 120 features + CatBoost.

2. nnU-Net v2: auto-configured 2D U-Net + identical features + XGBoost.

Hypotheses: (H1) nnU-Net v2 is superior in pixel segmentation (median DSC > 0.85); (H2) YOLOv8 is superior in speed and consistency (mAP > 98%, CV < 1%); (H3) features extracted from nnU-Net masks provide higher accuracy in sex/age prediction (AUC > 0.92); (H4) bilateral area difference provides the highest contribution in SHAP analysis (> 0.40). These thresholds were established a priori based on published benchmarks: DSC > 0.85 corresponds to “good” segmentation performance as reported in the Medical Segmentation Decathlon [22] and recommended by recent validation guidelines [54]; mAP > 98% aligns with YOLO-based dental segmentation benchmarks [21, 51]; AUC > 0.92 reflects the upper range of previously reported forensic sinus classification studies [11, 48, 60]; and SHAP > 0.40 was set to identify features with clinically meaningful contributions exceeding random baseline levels.