Department of Otolaryngology, Samsung Changwon Hospital, Sungkyunkwan University School of Medicine, Changwon 51353, Republic of Korea.
School of Electronic and Electrical Engineering, Hongik University, Seoul 04066, Republic of Korea.
Dentomaxillofac Radiol. 2024 Nov 1;53(8):549-557. doi: 10.1093/dmfr/twae042.
Sinusitis is a commonly encountered clinical condition that imposes a considerable burden on the healthcare systems. A significant number of maxillary sinus opacifications are diagnosed as sinusitis, often overlooking the precise differentiation between cystic formations and inflammatory sinusitis, resulting in inappropriate clinical treatment. This study aims to improve diagnostic accuracy by investigating the feasibility of differentiating maxillary sinusitis, retention cysts, and normal sinuses.
We developed a deep learning-based automatic detection model to diagnose maxillary sinusitis using ostiomeatal unit CT images. Of the 1080 randomly selected coronal-view CT images, including 2158 maxillary sinuses, datasets of maxillary sinus lesions comprised 1138 normal sinuses, 366 cysts, and 654 sinusitis based on radiographic findings, and were divided into training (n = 648 CT images), validation (n = 216), and test (n = 216) sets. We utilized a You Only Look Once based model for object detection, enhanced by the transfer learning method. To address the insufficiency of training data, various data augmentation techniques were adopted, thereby improving the model's robustness.
The trained You Only Look Once version 8 nano model achieved an overall precision of 97.1%, with the following class precisions on the test set: normal = 96.9%, cyst = 95.2%, and sinusitis = 99.2%. With an average F1-score of 95.4%, the F1-score was the highest for normal, then sinusitis, and finally, cysts. Upon evaluating a performance on difficulty level, the precision decreased to 92.4% on challenging test dataset.
The developed model is feasible for assisting clinicians in screening maxillary sinusitis lesions.
鼻窦炎是一种常见的临床病症，给医疗系统带来了相当大的负担。大量上颌窦混浊被诊断为鼻窦炎，往往忽略了对囊性形成和炎症性鼻窦炎的精确区分，导致临床治疗不当。本研究旨在通过研究区分上颌窦炎、潴留囊肿和正常窦腔的可行性来提高诊断准确性。
我们开发了一种基于深度学习的自动检测模型，使用中鼻甲-上颌窦复合体 CT 图像来诊断上颌窦炎。在随机选择的 1080 个冠状位 CT 图像中，包括 2158 个上颌窦，基于影像学发现，上颌窦病变数据集包括 1138 个正常窦、366 个囊肿和 654 个鼻窦炎，并分为训练集（n=648 个 CT 图像）、验证集（n=216 个）和测试集（n=216 个）。我们使用基于 You Only Look Once 的模型进行目标检测，并通过迁移学习方法进行增强。为了解决训练数据不足的问题，采用了各种数据增强技术，从而提高了模型的鲁棒性。
经过训练的 You Only Look Once 版本 8 纳米模型的整体精度达到了 97.1%，在测试集上的以下类别精度为：正常=96.9%，囊肿=95.2%，鼻窦炎=99.2%。平均 F1 得分为 95.4%，F1 得分最高的是正常，其次是鼻窦炎，最后是囊肿。在评估难度级别上的性能时，在具有挑战性的测试数据集上的精度降低到 92.4%。
所开发的模型可用于辅助临床医生筛选上颌窦炎病变。
