Abstract
Objectives: The differential diagnosis of ameloblastoma and odontogenic keratocyst is essential for surgical planning and patient counseling. While deep learning (DL)-based methods show promising potential in this domain, their clinical translation remains challenging due to insufficient interpretability. This study aims to introduce segmentation-guided preprocessing approaches to provide support for the clinical implementation of computer-aided diagnosis systems. Methods: This study evaluated the performance of an InceptionV3 model on 128 pathologically confirmed CBCT scans (AME: 64; OKC: 64) by 5-fold cross-validation. Four experimental inputs were compared: (1) Original slice; (2) Bounding-box ROI; (3) Precise segmentation ROI; and (4) Moderately expanded ROI. All models were trained under the same settings. Assessment was conducted on both the slice and patient levels, incorporating accuracy, recall, precision, F1-score, and the area under the receiver operating characteristic curve (AUC). Grad-CAM visualization and confidence curve analysis were employed to verify models' attention patterns and diagnostic confidence. Results: All models based on segmentation-guided ROI significantly outperformed models based on original slice. The moderately expanded ROI achieved optimal performance. The bounding-box ROI provided competitive performance with higher recall. Grad-CAM confirmed improved attention localization, while confidence curve analysis showed more consistent and reliable prediction patterns across slices. Conclusions: Segmentation-guided preprocessing represents an effective and clinically relevant approach for jaw lesion diagnosis and enhances interpretability.