Related Experiment Video
Updated: Jul 1, 2026
Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
Enhancing the preoperative differentiation of ameloblastoma and odontogenic keratocyst using pathomics-guided
Liang Li1, Shengfu Huang1, Kai Yang1
1State Key Laboratory of Oral & Maxillofacial Reconstruction and Regeneration, Key Laboratory of Oral Biomedicine Ministry of Education, Hubei Key Laboratory of Stomatology, School & Hospital of Stomatology, Wuhan University, Wuhan, China.
Background:
The cross-modal correlations between radiomics and pathomics, as well as their clinical translational applications in oral tumors and cysts, remain unclear. Here, we proposed a novel radiomic feature selection strategy guided by radio-pathomic correlations in Ameloblastoma (AM) and Odontogenic Keratocyst (OKC) to enhance their preoperative differentiation.
Methods:
We automatically extracted radiomic and pathomic features from CBCT scans and multi-resolution (25, 50, 100, 200 μm) whole slide image-derived cell density maps, respectively. Subsequently, radio-pathomic associations were evaluated by correlation analysis at two levels, directly between features, and between latent factors derived from features via factor analysis, separately in AM and OKC. Additionally, we compared the diagnostic performance of machine learning models using our proposed pathomics-guided feature selection strategy against traditional selection approaches.
Results:
At the feature level, correlation analysis identified one and seven significant feature pairs in AM and OKC, respectively (all |ρ| > 0.50, q < 0.05), suggesting that radiomic morphological feature weres strongly correlated with pathomic textural features reflecting tissue complexity. At the factor level, one significant factor pair was revealed in AM (ρ = 0.50, q = 0.007) and another in OKC (ρ = - 0.41, q = 0.04). Additionally, classification performance was enhanced by our proposed strategy across all six models, with an average area under the receiver operating characteristic curve (AUROC) improvement of 0.036 and individual gains ranging from 0.016 in Logistic Regression to 0.063 in Lasso.
Conclusion:
Significant cross-modal correlations between radiomic and pathomic features were identified in AM and OKC. Leveraging these associations, our proposed pathomics-guided radiomics showed the potential to improve the accuracy of preoperative differentiation between the two lesions, although these improvements reached statistical significance only in some models and still require further validation before clinical translation.