Automated skull orientation classification in Thai population using CNN-based transfer learning

Accurate classification of skull image orientations is a foundational prerequisite for advanced downstream applications, including automated sex estimation, ethnicity classification, and 3D cranial reconstruction. However, geometric variations and anatomical feature overlaps in cranial photographs present persistent challenges for computer vision models. This study introduces a confidence-based prediction rejection framework to establish a high-fidelity image preprocessing pipeline across five standard viewpoints: anterior (Front), posterior (Back), lateral (Side), anterolateral (Diagonal-Front), and posterolateral (Diagonal-Back). A dataset of 4,673 images derived from 43 dry skulls of Thai individuals was used to develop four convolutional neural network (CNN) models: ResNet50, VGG-16, MobileNetV2, and a Custom CNN. All models were evaluated using 20-fold cross-validation and validated on an independent test set of 760 images from 8 distinct skulls. VGG-16 achieved the highest cross-validation accuracy of 94.59% ±1.50% and lowest loss of 0.15 ± 0.04, whereas ResNet50 demonstrated superior generalization on the independent dataset. ResNet50 and the Custom CNN also exhibited stable and congruent convergence trajectories during optimization. A threshold-based filtering mechanism was deployed to mitigate false-positive errors, with τ > 0.90 identified as the optimal confidence cutoff. At this threshold, ResNet50 achieved an accuracy of 98.49% at a rejection rate of 29.87%. This framework serves as a reliable computational foundation for ensuring accurate skull view classification in forensic anthropology, craniofacial surgery, ethnicity classification, and sex estimation.