A bony projection arising from the temporal bone, termed the styloid process (SP), can undergo excessive lengthening and give rise to neck pain, throat irritation, and cephalalgia. This overgrowth, linked to Eagle syndrome, may exert pressure on adjacent nerves and vessels, occasionally triggering serious sequelae. The conventional imaging-driven categorization of elongated styloid process (ESP) variants is hampered by inconsistent radiographic quality, varying patient positioning, and individual anatomical variation—factors that, together, constrain diagnostic confidence. Modern strides in artificial intelligence, especially within deep learning, now permit a more robust classification workflow for ESP. The present work set out to engineer an automated ESP classification pipeline powered by deep learning models and to systematically compare the classification capabilities of two distinct network designs, EfficientNetB5 and InceptionV3. A retrospective design was adopted, with Orthopantomograms (OPGs) from our institutional oral radiology database used to classify ESP. The ImageJ application was used to measure the lengths of styloid processes. A curated collection of 330 elongated and 120 normal styloid process images formed the basis for model training and testing. Image preparation incorporated median filtering and dimensional resizing, and augmentation techniques were applied to strengthen generalisability. Both EfficientNetB5 and InceptionV3 served as feature-extraction backbones, each learning distinctive styloid-process representations. Model outputs were benchmarked on accuracy, precision, recall, and F1-score, and the resultant figures were compared side by side to determine which architecture offered greater clinical promise. The EfficientNetB5 model achieved 97.49% accuracy, 98.00% precision, 97.00% recall, and 97.00% F1-score, indicating a high level of discriminative power. Its AUC reached 0.9825. The InceptionV3 model achieved 84.11% accuracy, 85.00% precision, 84.00% recall, and an F1-score of 84.00%, with an AUC of 0.8943. Across all evaluated performance dimensions, EfficientNetB5 consistently exceeded InceptionV3. To summarise, this investigation delivers a deep learning–based solution, built on EfficientNetB5 and InceptionV3, for assigning elongated styloid processes to distinct morphological categories using digital panoramic radiographs. The evidence gathered suggests that these tools—above all, EfficientNetB5—can improve diagnostic accuracy and streamline clinical workflows, ultimately aiding the delivery of higher-quality patient care.
