Vered Marilena, Shnaiderman-Shapiro Anna, Malouf Rozet, Hirschhorn Ariel, Buchner Amos, Reiter Shoshana, Kats Lazar
Department of Oral Pathology, Oral Medicine and Maxillofacial Imaging, School of Dentistry, Tel Aviv University, 69978, Tel Aviv, Israel.
Institute of Pathology, Sheba Medical Center, Tel Hashomer, Israel.
Virchows Arch. 2025 Jun 27. doi: 10.1007/s00428-025-04160-z.
Microscopic images of aggressive and non-aggressive cases of central giant cell granuloma (CGCG) were analyzed by deep learning algorithms in order to assess its potential as a tool in predicting the biological behavior of CGCG. CGCGs with cortical expansion/perforation, tooth resorption/displacement, or recurrence were classified as aggressive (A-group; N = 48), CGCGs without these features as non-aggressive (N-group; N = 39). Data on patient age, gender, and jaw location were collected. Hematoxylin-eosin (H&E)-stained sections were scanned at × 10 magnification, yielding 9982 sections (5236 A-group, 4746 N-group). After excluding artifacts, 4272 sections (2629 A-group, 1643 N-group) were used to train a ResNet-50 model pre-trained on ImageNet. Data augmentation included random rotation, flipping, and zooming. Model was trained for 100 epochs with an 80/20 train/validation split and tested on 100 images (50 A-group, 50 N-group). Receiver Operating Characteristic (ROC) analysis with area under the curve (AUC), sensitivity, and specificity was performed; t-test and chi-square test were used for age and frequency (p < 0.05). AUC was 52%, sensitivity 54%, and specificity 50%. Mean age of patients in A-group was lower than in N-groups (32.6 ± 19.98 years and 42.2 ± 21.58 years, respectively; p = 0.038). F:M ratio was 1:1 in both groups. Mandible was twofold more frequently than maxilla in both groups. This pioneering study to differentiate between aggressive and non-aggressive CGCGs based on whole microscopic sections using a deep machine learning model was not successful, probably due to lack of specific segmentations and technical staining issues. Further investigation with advanced preprocessing is needed to enhance model performance and clinical utility.
为了评估深度学习算法在预测中央巨细胞肉芽肿（CGCG）生物学行为方面作为一种工具的潜力，对侵袭性和非侵袭性中央巨细胞肉芽肿病例的显微图像进行了分析。具有皮质扩张/穿孔、牙齿吸收/移位或复发的CGCG被分类为侵袭性（A组；N = 48），没有这些特征的CGCG被分类为非侵袭性（N组；N = 39）。收集了患者年龄、性别和颌骨位置的数据。苏木精-伊红（H&E）染色切片在×10放大倍数下扫描，共获得9982张切片（A组5236张，N组4746张）。排除伪像后，4272张切片（A组2629张，N组1643张）用于训练在ImageNet上预训练的ResNet-50模型。数据增强包括随机旋转、翻转和缩放。模型训练100个轮次，训练/验证分割比例为80/20，并在100张图像（A组50张，N组50张）上进行测试。进行了受试者操作特征（ROC）分析，计算曲线下面积（AUC）、敏感性和特异性；采用t检验和卡方检验分析年龄和频率（p < 0.05）。AUC为52%，敏感性为54%，特异性为50%。A组患者的平均年龄低于N组（分别为32.6±19.98岁和42.2±21.58岁；p = 0.038）。两组的女性与男性比例均为1:1。两组中下颌骨受累的频率均是上颌骨的两倍。这项基于整个显微切片使用深度机器学习模型区分侵袭性和非侵袭性CGCG的开创性研究未获成功，可能是由于缺乏特定的分割方法和技术染色问题。需要进一步采用先进的预处理方法进行研究，以提高模型性能和临床实用性。
