Methods: We retrospectively collected CT scan data from 276 patients with pathologically confirmed primary bone tumors from 4 medical centers in Guangdong Province between January, 2010 and August, 2021. A convolutional neural network (CNN) was employed as the deep learning architecture. The optimal baseline deep learning model (R-Net) was determined through transfer learning, and an optimized model (S-Net) was obtained through algorithmic improvements. Multivariate logistic regression analysis was used to screen the clinical features such as sex, age, mineralization location, and pathological fractures, which were then connected with the imaging features to construct the deep learning fusion model (SC-Net). The diagnostic performance of the SC-Net model and machine learning models were compared with radiologists' diagnoses, and their classification performance was evaluated using the area under the receiver operating characteristic curve (AUC) and F1 score.
Results: In the external test set, the fusion model (SC-Net) achieved the best performance with an AUC of 0.901 (95% CI: 0.803-1.00), an accuracy of 83.7% (95% CI: 69.3%-93.2%) and an F1 score of 0.857, and outperformed the S-Net model with an AUC of 0.818 (95% CI: 0.694-0.942), an accuracy of 76.7% (95% CI: 61.4%-88.2%), and an F1 score of 0.828. The overall classification performance of the fusion model (SC-Net) exceeded that of radiologists' diagnoses.
Conclusions: The deep learning fusion model based on multi-center CT images and clinical features is capable of accurate classification of osseous and chondroid matrix mineralization and may potentially improve the accuracy of clinical diagnoses of osteogenic versus chondrogenic primary bone tumors.
目的: 基于CT图像和临床特征构建深度学习模型，鉴别原发性骨肿瘤骨样和软骨样基质矿化，以提示成骨性或成软骨性骨肿瘤的组织来源，辅助两者的鉴别诊断。方法: 回顾性搜集2010年1月~2021年8月来自广东省4个医疗中心的276例病理证实的原发性骨肿瘤患者CT平扫图像。采用卷积神经网络（CNN）作为深度学习架构，通过迁移学习确定最佳深度学习基线模型（R-Net），通过算法改进获得优化后的深度学习模型（S-Net），采用多元逻辑回归分析筛选性别、年龄、矿化位置和病理性骨折等临床特征，将临床特征与影像特征连接构建深度学习融合模型（SC-Net）。对比深度学习模型与机器学习模型、放射科医生的诊断表现。用受试者特征曲线（ROC）下面积（AUC）和F1分数评价模型分类性能。结果: 外部测试集显示：深度学习融合模型SC-Net的表现最佳，AUC为0.901（95% CI：0.803~1.00），准确度为83.7%（95% CI：69.3%~93.2%），F1分数为0.857，性能优于深度学习模型R-Net、深度学习模型S-Net、机器学习模型和机器学习融合模型，AUC分别为0.768、0.818、0.761、0.791，准确度为69.8%、76.7%、72.1%、74.4%，F1分数为0.755、0.828、0.700、0.732；且深度学习融合模型SC-Net总体分类性能超越了放射科医生诊断水平。结论: 基于多中心的CT图像和临床信息的深度学习融合模型，成功实现了对原发性骨肿瘤骨样和软骨样基质矿化的分类。尤其对于影像表现不典型矿化病灶的鉴别优于机器学习模型和放射科医生视觉诊断，具有一定的临床应用价值。.
Keywords: chondroid matrix mineralization; computed tomography; deep learning; osteoid matrix mineralization; primary bone tumor.