Deep Learning on Histology Images for Differentiating Fibro-osseous Lesions of the Jaw

Abstract

Fibrous dysplasia (FD), cemento-ossifying fibroma (COF), and cemento-osseous dysplasia (COD) are fibro-osseous lesions of the jaw that share histologic features, making their differentiation in routine pathology practice challenging. Distinguishing these conditions is crucial for clinical management due to their varying prognoses. To address these limitations, we developed and validated a deep learning model using 1,218 hematoxylin and eosin whole slide images from 338 cases (FD, n = 130; COF, n = 146; COD, n = 62) across 3 institutions. After digital scanning and image preprocessing, we compared 4 training strategies using a ResNet-50 backbone with loss functions and multiple-instance learning: weakly and fully supervised models on single or multiple slides. In the test set, the weakly supervised model on multislide achieved the best performance (area under the curve, 0.86; accuracy, 0.71) as compared with other models, exceeding the diagnostic accuracy of experienced oral pathologists. Visualization with probability heat maps revealed that the model identified key histomorphologic patterns critical for distinguishing FD, COF, and COD. These results demonstrate that a multislide, weakly supervised deep learning approach can provide a classification aid for fibro-osseous lesions with higher discriminative performance than oral pathologists when only histologic slides are considered, highlighting its potential as a supportive tool alongside clinical, radiologic, and molecular data. However, we acknowledge that the test cohort is limited in sample size and geographic diversity; thus, while the results are promising, further expansion and more diverse cohorts would be needed to fully substantiate the generalizability of the model.