Automated Biological Sex Estimation from Two-Dimensional Dental Arch Images Using Deep Learning

<title>Abstract</title> <p>Sex estimation is a foundational step in forensic identification, yet existing automated dental methods depend on ionising radiography — orthopantomography (OPG) or cone-beam computed tomography (CBCT) — which imposes radiation exposure, capital equipment requirements and deployment barri ers that limit use in disaster victim identification (DVI). We evaluated whether biological sex can be estimated end-to-end from radiation-free two-dimensional occlusal projections of the dental arch. A bal anced corpus of 6,000 images (3,000 male, 3,000 female; aged 18–65 years) was derived from digital intraoral scans acquired at a single institution and orthogonally projected onto the occlusal plane. A held out test set of 1,000 images was reserved before any model fitting, and the remaining 5,000 images were partitioned into five stratified cross-validation folds. Four convolutional architectures were benchmarked under an identical protocol: a custom baseline, EfficientNetB0, MobileNetV2 and ResNet50. Efficient NetB0 achieved the highest accuracy (96.39 ± 0.84%; 95% CI 95.46–97.32%; AUC 0.980 ± 0.006), significantly exceeding the custom baseline and ResNet50 (Bonferroni-corrected paired t-tests, p &lt; 0.001) but not MobileNetV2 (95.88 ± 0.91%; p = 0.042). A factorial ablation showed that contrast limited adaptive histogram equalisation and geometric augmentation contributed a near-additive 4.69 percentage-point gain. Qualitative inspection of gradient-weighted class activation maps indicated that model attention concentrated on the inter-canine and inter-molar regions, the sites identified by popula tion morphometrics as maximally dimorphic. Dental arch geometry is therefore a viable radiation-free modality for forensic sex estimation, with MobileNetV2 preferable for memory-constrained field de ployment.</p>