Rašić Mario, Tropčić Mario, Pupić-Bakrač Jure, Subašić Marko, Čvrljević Igor, Dediol Emil
Clinic for Tumors, Clinical Hospital Center "Sisters of Mercy", Ilica 197, 10000 Zagreb, Croatia.
Faculty of Electrical Engineering and Computing, University of Zagreb, Unska ulica 3, 10000 Zagreb, Croatia.
Diagnostics (Basel). 2024 Jul 6;14(13):1443. doi: 10.3390/diagnostics14131443.
The purpose of this study was to develop a deep learning algorithm capable of diagnosing radicular cysts in the lower jaw on panoramic radiographs.
In this study, we conducted a comprehensive analysis of 138 radicular cysts and 100 normal panoramic radiographs collected from 2013 to 2023 at Clinical Hospital Dubrava. The images were annotated by a team comprising a radiologist and a maxillofacial surgeon, utilizing the GNU Image Manipulation Program. Furthermore, the dataset was enriched through the application of various augmentation techniques to improve its robustness. The evaluation of the algorithm's performance and a deep dive into its mechanics were achieved using performance metrics and EigenCAM maps.
In the task of diagnosing radicular cysts, the initial algorithm performance-without the use of augmentation techniques-yielded the following scores: precision at 85.8%, recall at 66.7%, mean average precision (mAP)@50 threshold at 70.9%, and mAP@50-95 thresholds at 60.2%. The introduction of image augmentation techniques led to the precision of 74%, recall of 77.8%, mAP@50 threshold to 89.6%, and mAP@50-95 thresholds of 71.7, respectively. Also, the precision and recall were transformed into F1 scores to provide a balanced evaluation of model performance. The weighted function of these metrics determined the overall efficacy of our models. In our evaluation, non-augmented data achieved F1 scores of 0.750, while augmented data achieved slightly higher scores of 0.758.
Our study underscores the pivotal role that deep learning is poised to play in the future of oral and maxillofacial radiology. Furthermore, the algorithm developed through this research demonstrates a capability to diagnose radicular cysts accurately, heralding a significant advancement in the field.
本研究的目的是开发一种能够在全景X线片上诊断下颌根端囊肿的深度学习算法。
在本研究中，我们对2013年至2023年在杜布拉瓦临床医院收集的138例根端囊肿和100例正常全景X线片进行了综合分析。这些图像由一名放射科医生和一名颌面外科医生组成的团队使用GNU图像处理程序进行标注。此外，通过应用各种增强技术丰富数据集，以提高其鲁棒性。使用性能指标和特征激活映射图对算法性能进行评估并深入研究其机制。
在诊断根端囊肿的任务中，初始算法性能（未使用增强技术）得出以下分数：精度为85.8%，召回率为66.7%，50阈值下的平均精度均值（mAP）为70.9%，50 - 95阈值下的mAP为60.2%。图像增强技术的引入分别使精度达到74%，召回率达到77.8%，50阈值下的mAP达到89.6%，50 - 95阈值下的mAP达到71.7%。此外，将精度和召回率转换为F1分数，以对模型性能进行平衡评估。这些指标的加权函数决定了我们模型的整体效能。在我们的评估中，未增强数据的F1分数为0.750，而增强数据的分数略高，为0.758。
我们的研究强调了深度学习在口腔颌面放射学未来将发挥的关键作用。此外，通过本研究开发的算法显示出能够准确诊断根端囊肿的能力，预示着该领域的重大进展。
