:刘木清, 傅开元. 深度学习辅助口腔颌面医学影像学的研究现状与展望[J]. 中华口腔医学杂志, 2023, 58(6): 533-539. DOI: 10.3760/cma.j.cn112144-20230405-00138.利益冲突:所有作者声明不存在利益冲突参考文献
[1] 唐莉, 徐欣, 廖玍. 口腔医学领域人工智能研究前沿和热点分析[J]. 中国循证医学杂志, 2022, 22(6): 684-691. DOI: 10.7507/1672-2531.202202103.Tang L, Xu X, Liao G. Analysis of frontiers and hotspots of artificial intelligence applied in stomatology[J]. Chin J Evid Based Med, 2022, 22(6): 684-691. DOI: 10.7507/1672-2531.202202103.
[2] Hung KF, Yeung A, Bornstein MM, et al. Personalized dental medicine, artificial intelligence, and their relevance for dentomaxillofacial imaging[J]. Dentomaxillofac Radiol, 2023, 52(1): 20220335. DOI: 10.1259/dmfr.20220335.
[3] Chen H, Zhang K, Lyu P, et al. A deep learning approach to automatic teeth detection and numbering based on object detection in dental periapical films[J]. Sci Rep, 2019, 9(1): 3840. DOI: 10.1038/s41598-019-40414-y.
[4] Tuzoff DV, Tuzova LN, Bornstein MM, et al. Tooth detection and numbering in panoramic radiographs using convolutional neural networks[J]. Dentomaxillofac Radiol, 2019, 48(4): 20180051. DOI: 10.1259/dmfr.20180051.
[5] K?l?c MC, Bayrakdar IS, ?elik ?, et al. Artificial intelligence system for automatic deciduous tooth detection and numbering in panoramic radiographs[J]. Dentomaxillofac Radiol, 2021, 50(6): 20200172. DOI: 10.1259/dmfr.20200172.
[6] Lee JH, Han SS, Kim YH, et al. Application of a fully deep convolutional neural network to the automation of tooth segmentation on panoramic radiographs[J]. Oral Surg Oral Med Oral Pathol Oral Radiol, 2020, 129(6): 635-642. DOI: 10.1016/j.oooo.2019.11.007.
[7] Wang H, Minnema J, Batenburg KJ, et al. Multiclass CBCT image segmentation for orthodontics with deep learning[J]. J Dent Res, 2021, 100(9): 943-949. DOI: 10.1177/00220345211005338.
[8] Chen YL, Du HY, Yun ZQ, et al. Automatic segmentation of individual tooth in dental CBCT images from tooth surface map by a multi-task FCN[J]. IEEE Access, 2020, 8: 97296-97309. DOI: 10.1109/ACCESS.2020.2991799.
[9] Shaheen E, Leite A, Alqahtani KA, et al. A novel deep learning system for multi-class tooth segmentation and classification on cone beam computed tomography. A validation study[J]. J Dent, 2021, 115: 103865. DOI: 10.1016/j.jdent.2021.103865.
[10] Cui Z, Fang Y, Mei L, et al. A fully automatic AI system for tooth and alveolar bone segmentation from cone-beam CT images[J]. Nat Commun, 2022, 13(1): 2096. DOI: 10.1038/s41467-022-29637-2.
[11] Lahoud P, Diels S, Niclaes L, et al. Development and validation of a novel artificial intelligence driven tool for accurate mandibular canal segmentation on CBCT[J]. J Dent, 2022, 116: 103891. DOI: 10.1016/j.jdent.2021.103891.
[12] Liu MQ, Xu ZN, Mao WY, et al. Deep learning-based evaluation of the relationship between mandibular third molar and mandibular canal on CBCT[J]. Clin Oral Investig, 2022, 26(1): 981-991. DOI: 10.1007/s00784-021-04082-5.
[13] Bayraktar Y, Ayan E. Diagnosis of interproximal caries lesions with deep convolutional neural network in digital bitewing radiographs[J]. Clin Oral Investig, 2022, 26(1): 623-632. DOI: 10.1007/s00784-021-04040-1.
[14] Chen H, Li H, Zhao Y, et al. Dental disease detection on periapical radiographs based on deep convolutional neural networks[J]. Int J Comput Assist Radiol Surg, 2021, 16(4): 649-661. DOI: 10.1007/s11548-021-02319-y.
[15] Lee S, Oh SI, Jo J, et al. Deep learning for early dental caries detection in bitewing radiographs[J]. Sci Rep, 2021, 11(1): 16807. DOI: 10.1038/s41598-021-96368-7.
[16] Lian L, Zhu T, Zhu F, et al. Deep learning for caries detection and classification[J]. Diagnostics (Basel), 2021, 11(9): 1672. DOI: 10.3390/diagnostics11091672.
[17] Mohammad-Rahimi H, Motamedian SR, Rohban MH, et al. Deep learning for caries detection: a systematic review[J]. J Dent, 2022, 122: 104115. DOI: 10.1016/j.jdent.2022.104115.
[18] Zheng ZY, Yan H, Setzer FC, et al. Anatomically constrained deep learning for automating dental CBCT segmentation and lesion detection[J]. IEEE Trans Autom Sci Eng, 2021, 18(2): 603-614. DOI: 10.1109/TASE.2020.3025871.
[19] 钱军, 马芮, 曲妍, 等. 人工智能在锥形束计算机断层扫描影像中识别慢性根尖周炎根尖区病变的应用[J]. 华西口腔医学杂志, 2022, 40(5): 576-581. DOI: 10.7518/hxkq.2022.05.011.Qian J, Ma R, Qu Y, et al. Use and performance of artificial intelligence applications in the diagnosis of chronic apical periodontitis based on cone beam computed tomography imaging[J]. West China J Stomatol, 2022, 40(5): 576-581. DOI: 10.7518/hxkq.2022.05.011.
[20] Yu D, Hu J, Feng Z, et al. Deep learning based diagnosis for cysts and tumors of jaw with massive healthy samples[J]. Sci Rep, 2022, 12(1): 1855. DOI: 10.1038/s41598-022-05913-5.
[21] Kwon O, Yong TH, Kang SR, et al. Automatic diagnosis for cysts and tumors of both jaws on panoramic radiographs using a deep convolution neural network[J]. Dentomaxillofac Radiol, 2020, 49(8): 20200185. DOI: 10.1259/dmfr.20200185.
[22] Lee JH, Kim DH, Jeong SN. Diagnosis of cystic lesions using panoramic and cone beam computed tomographic images based on deep learning neural network[J]. Oral Dis, 2020, 26(1): 152-158. DOI: 10.1111/odi.13223.
[23] Xu SP, Liu C, Zong YS, et al. An early diagnosis of oral cancer based on three-dimensional convolutional neural networks[J]. IEEE Access, 2019, 7: 158603-158611. DOI: 10.1109/ACCESS.2019.2950286.
[24] Schouten J, Noteboom S, Martens RM, et al. Automatic segmentation of head and neck primary tumors on MRI using a multi-view CNN[J]. Cancer Imaging, 2022, 22(1): 8. DOI: 10.1186/s40644-022-00445-7.
[25] Ariji Y, Fukuda M, Kise Y, et al. Contrast-enhanced computed tomography image assessment of cervical lymph node metastasis in patients with oral cancer by using a deep learning system of artificial intelligence[J]. Oral Surg Oral Med Oral Pathol Oral Radiol, 2019, 127(5): 458-463. DOI: 10.1016/j.oooo.2018.10.002.
[26] Ariji Y, Sugita Y, Nagao T, et al. CT evaluation of extranodal extension of cervical lymph node metastases in patients with oral squamous cell carcinoma using deep learning classification[J]. Oral Radiol, 2020, 36(2): 148-155. DOI: 10.1007/s11282-019-00391-4.
[27] Ariji Y, Fukuda M, Nozawa M, et al. Automatic detection of cervical lymph nodes in patients with oral squamous cell carcinoma using a deep learning technique: a preliminary study[J]. Oral Radiol, 2021, 37(2): 290-296. DOI: 10.1007/s11282-020-00449-8.
[28] Ariji Y, Kise Y, Fukuda M, et al. Segmentation of metastatic cervical lymph nodes from CT images of oral cancers using deep-learning technology[J]. Dentomaxillofac Radiol, 2022, 51(4): 20210515. DOI: 10.1259/dmfr.20210515.
[29] Xu X, Xi L, Wei L, et al. Deep learning assisted contrast-enhanced CT-based diagnosis of cervical lymph node metastasis of oral cancer: a retrospective study of 1466 cases[J]. Eur Radiol, 2023, 33(6): 4303-4312. DOI: 10.1007/s00330-022-09355-5.
[30] Fujima N, Andreu-Arasa VC, Meibom SK, et al. Deep learning analysis using FDG-PET to predict treatment outcome in patients with oral cavity squamous cell carcinoma[J]. Eur Radiol, 2020, 30(11): 6322-6330. DOI: 10.1007/s00330-020-06982-8.
[31] Ariji Y, Mori M, Fukuda M, et al. Automatic visualization of the mandibular canal in relation to an impacted mandibular third molar on panoramic radiographs using deep learning segmentation and transfer learning techniques[J]. Oral Surg Oral Med Oral Pathol Oral Radiol, 2022, 134(6): 749-757. DOI: 10.1016/j.oooo.2022.05.014.
[32] Zhu T, Chen D, Wu F, et al. Artificial intelligence model to detect real contact relationship between mandibular third molars and inferior alveolar nerve based on panoramic radiographs[J]. Diagnostics (Basel), 2021, 11(9): 1664. DOI: 10.3390/diagnostics11091664.
[33] Orhan K, Bilgir E, Bayrakdar IS, et al. Evaluation of artificial intelligence for detecting impacted third molars on cone-beam computed tomography scans[J]. J Stomatol Oral Maxillofac Surg, 2021, 122(4): 333-337. DOI: 10.1016/j.jormas.2020.12.006.
[34] Park JH, Hwang HW, Moon JH, et al. Automated identification of cephalometric landmarks: part 1-comparisons between the latest deep-learning methods YOLOV3 and SSD[J]. Angle Orthod, 2019, 89(6): 903-909. DOI: 10.2319/022019-127.1.
[35] Hwang HW, Park JH, Moon JH, et al. Automated identification of cephalometric landmarks: part 2-might it be better than human?[J]. Angle Orthod, 2020, 90(1): 69-76. DOI: 10.2319/022019-129.1.
[36] Chen X, Lian C, Deng HH, et al. Fast and accurate craniomaxillofacial landmark detection via 3D Faster R-CNN[J]. IEEE Trans Med Imaging, 2021, 40(12): 3867-3878. DOI: 10.1109/TMI.2021.3099509.
[37] Lang Y, Lian C, Xiao D, et al. Localization of craniomaxillofacial landmarks on CBCT images using 3D Mask R-CNN and local dependency learning[J]. IEEE Trans Med Imaging, 2022, 41(10): 2856-2866. DOI: 10.1109/TMI.2022.3174513.
[38] Yu HJ, Cho SR, Kim MJ, et al. Automated skeletal classification with lateral cephalometry based on artificial intelligence[J]. J Dent Res, 2020, 99(3): 249-256. DOI: 10.1177/0022034520901715.
[39] Mahto RK, Kafle D, Giri A, et al. Evaluation of fully automated cephalometric measurements obtained from web-based artificial intelligence driven platform[J]. BMC Oral Health, 2022, 22(1): 132. DOI: 10.1186/s12903-022-02170-w.
[40] 朱玉佳, 许晴, 赵一姣, 等. 深度学习算法辅助构建三维颜面正中矢状平面[J]. 北京大学学报(医学版), 2022, 54(1): 134-139. DOI: 10.19723/j.issn.1671-167X.2022.01.021.Zhu YJ, Xu Q, Zhao YJ, et al. Deep learning-assisted construction of three-dimensional facial midsagittal plane[J]. J Peking Univ (Health Sci), 2022, 54(1): 134-139. DOI: 10.19723/j.issn.1671-167X.2022.01.021.
[41] Zhou J, Zhou H, Pu L, et al. Development of an artificial intelligence system for the automatic evaluation of cervical vertebral maturation status[J]. Diagnostics (Basel), 2021, 11(12): 2200. DOI: 10.3390/diagnostics11122200.
[42] Suhail Y, Upadhyay M, Chhibber A, et al. Machine learning for the diagnosis of orthodontic extractions: a computational analysis using ensemble learning[J]. Bioengineering (Basel), 2020, 7(2): 55. DOI: 10.3390/bioengineering7020055.
[43] Lee JH, Kim DH, Jeong SN, et al. Diagnosis and prediction of periodontally compromised teeth using a deep learning-based convolutional neural network algorithm[J]. J Periodontal Implant Sci, 2018, 48(2): 114-123. DOI: 10.5051/jpis.2018.48.2.114.
[44] Kim J, Lee HS, Song IS, et al. DeNTNet: deep neural transfer network for the detection of periodontal bone loss using panoramic dental radiographs[J]. Sci Rep, 2019, 9(1): 17615. DOI: 10.1038/s41598-019-53758-2.
[45] Chang HJ, Lee SJ, Yong TH, et al. Deep learning hybrid method to automatically diagnose periodontal bone loss and stage periodontitis[J]. Sci Rep, 2020, 10(1): 7531. DOI: 10.1038/s41598-020-64509-z.
[46] Kurt Bayrakdar S, Orhan K, Bayrakdar IS, et al. A deep learning approach for dental implant planning in cone-beam computed tomography images[J]. BMC Med Imaging, 2021, 21(1): 86. DOI: 10.1186/s12880-021-00618-z.
[47] Hadj Sa?d M, Le Roux MK, Catherine JH, et al. Development of an artificial intelligence model to identify a dental implant from a radiograph[J]. Int J Oral Maxillofac Implants, 2020, 36(6): 1077-1082. DOI: 10.11607/jomi.8060.
[48] Sukegawa S, Yoshii K, Hara T, et al. Is attention branch network effective in classifying dental implants from panoramic radiograph images by deep learning?[J]. PLoS One, 2022, 17(7): e0269016. DOI: 10.1371/journal.pone.0269016.
[49] Liu M, Wang S, Chen H, et al. A pilot study of a deep learning approach to detect marginal bone loss around implants[J]. BMC Oral Health, 2022, 22(1): 11. DOI: 10.1186/s12903-021-02035-8.
[50] Lee DW, Kim SY, Jeong SN, et al. Artificial intelligence in fractured dental implant detection and classification: evaluation using dataset from two dental hospitals[J]. Diagnostics (Basel), 2021, 11(2): 233. DOI: 10.3390/diagnostics11020233.
[51] Lee JH, Kim YT, Lee JB, et al. A performance comparison between automated deep learning and dental professionals in classification of dental implant systems from dental imaging: a multi-center study[J]. Diagnostics (Basel), 2020, 10(11): 910. DOI: 10.3390/diagnostics10110910.
[52] Wang X, Xu Z, Tong Y, et al. Detection and classification of mandibular fracture on CT scan using deep convolutional neural network[J]. Clin Oral Investig, 2022, 26(6): 4593-4601. DOI: 10.1007/s00784-022-04427-8.
[53] Kuwana R, Ariji Y, Fukuda M, et al. Performance of deep learning object detection technology in the detection and diagnosis of maxillary sinus lesions on panoramic radiographs[J]. Dentomaxillofac Radiol, 2021, 50(1): 20200171. DOI: 10.1259/dmfr.20200171.
[54] Lee JS, Adhikari S, Liu L, et al. Osteoporosis detection in panoramic radiographs using a deep convolutional neural network-based computer-assisted diagnosis system: a preliminary study[J]. Dentomaxillofac Radiol, 2019, 48(1): 20170344. DOI: 10.1259/dmfr.20170344.
[55] Kise Y, Ikeda H, Fujii T, et al. Preliminary study on the application of deep learning system to diagnosis of Sj?gren′s syndrome on CT images[J]. Dentomaxillofac Radiol, 2019, 48(6): 20190019. DOI: 10.1259/dmfr.20190019.
[56] Choi E, Kim D, Lee JY, et al. Artificial intelligence in detecting temporomandibular joint osteoarthritis on orthopantomogram[J]. Sci Rep, 2021, 11(1): 10246. DOI: 10.1038/s41598-021-89742-y.
[57] Lee KS, Kwak HJ, Oh JM, et al. Automated detection of TMJ osteoarthritis based on artificial intelligence[J]. J Dent Res, 2020, 99(12): 1363-1367. DOI: 10.1177/0022034520936950.
[58] Lee YH, Won JH, Kim S, et al. Advantages of deep learning with convolutional neural network in detecting disc displacement of the temporomandibular joint in magnetic resonance imaging[J]. Sci Rep, 2022, 12(1): 11352. DOI: 10.1038/s41598-022-15231-5.
[59] Merdietio Boedi R, Banar N, De Tobel J, et al. Effect of lower third molar segmentations on automated tooth development staging using a convolutional neural network[J]. J Forensic Sci, 2020, 65(2): 481-486. DOI: 10.1111/1556-4029.14182.
[60] Mu CC, Li G. Age estimation using panoramic radiographs by transfer learning[J]. Chin J Dent Res, 2022, 25(2): 119-124. DOI: 10.3290/j.cjdr.b3086341.
[61] Zheng Q, Ge Z, Du H, et al. Age estimation based on 3D pulp chamber segmentation of first molars from cone-beam-computed tomography by integrated deep learning and level set[J]. Int J Legal Med, 2021, 135(1): 365-373. DOI: 10.1007/s00414-020-02459-x.
[62] Patil V, Vineetha R, Vatsa S, et al. Artificial neural network for gender determination using mandibular morphometric parameters: a comparative retrospective study[J]. Cogent Eng, 2020, 7(1): 1723783. DOI: 10.1080/23311916.2020.1723783.
[63] Nishiyama M, Ishibashi K, Ariji Y, et al. Performance of deep learning models constructed using panoramic radiographs from two hospitals to diagnose fractures of the mandibular condyle[J]. Dentomaxillofac Radiol, 2021, 50(7): 20200611. DOI: 10.1259/dmfr.20200611.
[64] Krois J, Garcia Cantu A, Chaurasia A, et al. Generalizability of deep learning models for dental image analysis[J]. Sci Rep, 2021, 11(1): 6102. DOI: 10.1038/s41598-021-85454-5.
[65] Finlayson SG, Bowers JD, Ito J, et al. Adversarial attacks on medical machine learning[J]. Science, 2019, 363(6433): 1287-1289. DOI: 10.1126/science.aaw4399.
[66] Schwendicke F, Singh T, Lee JH, et al. Artificial intelligence in dental research: checklist for authors, reviewers, readers[J]. J Dent, 2021, 107: 103610. DOI: 10.1016/j.jdent.2021.103610.
[67] Mongan J, Moy L, Kahn CEJr. Checklist for artificial intelligence in medical imaging (CLAIM): a guide for authors and reviewers[J]. Radiol Artif Intell, 2020, 2(2): e200029. DOI: 10.1148/ryai.2020200029.
[68] Rieke N, Hancox J, Li W, et al. The future of digital health with federated learning[J]. NPJ Digit Med, 2020, 3: 119. DOI: 10.1038/s41746-020-00323-1.
[69] Zhang Y, Hong D, McClement D, et al. Grad-CAM helps interpret the deep learning models trained to classify multiple sclerosis types using clinical brain magnetic resonance imaging[J]. J Neurosci Methods, 2021, 353: 109098. DOI: 10.1016/j.jneumeth.2021.109098.
[70] Rodriguez D, Nayak T, Chen Y, et al. On the role of deep learning model complexity in adversarial robustness for medical images[J]. BMC Med Inform Decis Mak, 2022, 22(Suppl 2): 160. DOI: 10.1186/s12911-022-01891-w.
近年,以深度学习为代表的人工智能在医疗领域飞速发展,取得令人瞩目的成果。口腔颌面医学影像,如口内X线片(根尖片、 翼片等)、口外X线片(曲面体层X线片、头颅侧位X线片等)、锥形束CT、螺旋CT及MRI等,是口腔疾病诊断和治疗不可或缺的部分,是口腔医学领域深度学习研究的主要数据[1]。深度学习在口腔颌面影像中的主要应用包括图像分析与影像质量提升。目前相关研究多处于临床应用前阶段,并逐步走向临床应用[2]。本文主要阐述深度学习用于口腔颌面医学影像图像分析方面的研究现状,并分析目前研究的局限以及展望未来的研究方向。
深度学习在口腔颌面医学影像分析中的主要应用包括:①牙及颌面部其他解剖结构的检测、识别和分割等;②口腔颌面部疾病的检测与诊断;③法医学牙齿鉴定。卷积神经网络(convolutional neural network,CNN)是目前医学图像深度学习研究应用最多的算法。深度学习研究中检测及分类任务常用评价指标包括:准确度、召回率(也称灵敏度)、特异度、精确率(也称阳性预测值)、F1分数,以及受试者操作特征曲线、PR曲线(precision-recall curve)等。分割是划定图像中兴趣区域的轮廓,是像素级别的检测,常用评价指标包括:准确度、Dice相似系数(Dice similarity coefficient,DSC)、交并比、豪斯多夫距离、平均表面对称距离等。DSC及交并比取值范围为0~1,越接近1说明分割结果越准确。
1.牙及颌面部其他解剖结构的检测、识别和分割:在二维影像如根尖片与曲面体层X线片解剖结构的检测和识别方面,Chen等[3]收集1 250张根尖片,采用Faster R-CNN模型建立牙齿自动检测和识别方法,精确率和召回率均超过0.9。曲面体层X线片作为最常用的口外X线片,可显示全牙列及颌骨情况。Tuzoff等[4]及K?l?c等[5]采用Faster R-CNN模型分别实现曲面体层X线片恒牙列及乳牙列牙齿的准确检测和识别。在曲面体层X线片牙齿分割方面,Lee等[6]证实基于CNN的自动分割模型交并比可达0.877。
在三维影像解剖结构检测、识别与分割方面,锥形束CT解剖结构的自动准确识别有助于数字化口腔诊疗的开展。牙齿的自动分割可用于疾病诊断及正畸、修复中数字化模拟和设计。颌骨的分割及解剖标志点的识别可用于头影测量、手术设计等。颌骨中精细解剖结构的分割也有重要意义,如上颌窦、下颌管的分割可用于种植、智齿拔除及其他外科手术的风险评价。多数锥形束CT深度学习相关研究基于数十例至百余例的小样本数据[7, 8, 9]。Cui等[10]建立了目前最大的锥形束CT数据集,包括15个中心的4 938例锥形束CT,利用牙形态学信息结合大样本数据训练实现了颌骨及牙的准确分割及识别,DSC分别为0.930和0.915,与经验丰富的影像专业医师相当。还有研究报道了颌骨精细结构如下颌管的自动检测[11, 12]。笔者课题组采用基于U-net的深度学习模型,实现小视野锥形束CT下颌管磨牙段的准确分割,DSC可达0.92[12]。
2.口腔颌面部疾病检测与辅助诊断:口腔颌面部疾病诊断方面,深度学习在牙体牙髓、口腔颌面外科、正畸、牙周、种植等多个学科的疾病诊断中显示出光明的前景,部分人工智能模型可达到甚至超过相关领域专家的水平。
(1)龋病及根尖周病变:X线片龋及慢性根尖周炎的诊断具有挑战性,特别是早期病变。目前已有较多研究报道X线片龋病的自动检测、分类与病变分割[13, 14, 15, 16]。2022年的一项系统综述显示,龋分类模型的准确度存在差异,根尖片为82%~99.2%, 翼片为87.6%~95.4%,曲面体层X线片为86.1%~96.1%[17]。Zheng等[18]进行小视野锥形束CT慢性根尖周炎的自动检测研究,精确率为0.9,召回率为0.84。钱军等[19]采用3D U-net模型在锥形束CT中检测及分割慢性根尖周炎病变,DSC可达0.959。
(2)口腔颌面部肿瘤:在颌骨囊肿与良性肿瘤方面,曲面体层X线片是颌骨肿瘤的首选影像学检查方法。多种类型CNN模型,如GoogLeNet Inception、YOLO等被用于颌骨囊肿及肿瘤的诊断研究[20, 21, 22],其中部分研究报道诊断表现与口腔颌面外科专业医师相当。Lee等[22]收集曲面体层X线片及锥形束CT数据建立基于Inception v3的模型,可有效检测和鉴别常见的3种牙源性颌骨囊肿,并发现基于锥形束CT的诊断模型表现明显优于曲面体层模型。
在口腔鳞状细胞癌方面,Xu等[23]建立了一种基于3D CNN模型的增强CT早期口腔癌诊断系统,结合动态增强信息,曲线下面积(area under the curve,AUC)可达0.8。Schouten等[24]使用多视图CNN在MRI中自动分割头颈部鳞状细胞癌病灶,平均DSC为0.49;肿瘤体积越大,分割越准确。口腔鳞状细胞癌易发生颈部淋巴结转移。Ariji等[25, 26, 27, 28]开展了一系列研究,从人工裁剪增强CT淋巴结的图像诊断是否存在转移及包膜外侵犯,到从图像中自动检测和分割淋巴结;最新的模型检测颈淋巴结转移AUC可达0.95,超过2名从业超过20年的口腔颌面影像专业医师[28]。Xu等[29]开发了基于1 466例口腔鳞状细胞癌患者增强CT数据的深度学习模型,用于淋巴结定位以及判断是否存在转移,对淋巴结转移诊断的准确度与放射科医师相似,远高于外科医师及医学生。肿瘤预后方面,Fujima等[30]使用正电子发射体层成像和CT图像开发基于深度学习的口腔癌患者预后预测模型,与传统T分期、临床分期相比,模型可更准确预测患者的无病生存率。
(3)阻生智齿:深度学习也被用于阻生智齿的自动诊断,研究大多基于曲面体层X线片[31, 32],锥形束CT研究相对较少[12,33]。Orhan等[33]对人工智能系统(Diagnocat)用于锥形束CT阻生智齿的诊断性能进行评估,对下颌智齿与下颌管关系的判断与医师结果的一致性Kappa值为0.762。笔者课题组结合基于U-Net的下颌智齿和下颌管自动分割模型与基于ResNet-34的分类模型,对下颌智齿与下颌管的空间关系进行分类,结果灵敏度为0.90,特异度为0.95,与口腔影像专业住院医师水平相当[12]。
(4)正畸:目前已有多项研究应用深度学习进行头颅侧位X线片及锥形束CT头影测量标志点的自动定位及测量[34, 35, 36, 37],或直接对颌骨进行分类诊断[38]。2020年,Yu等[38]建立基于5 890张头颅侧位X线片的CNN诊断系统,颌骨分类准确度可达0.964。目前已有商业软件应用于正畸自动头影测量,准确度与医师手工测量相当[39]。朱玉佳等[40]将深度学习与赋权普氏分析算法结合,用于三维颜面正中矢状平面的自动构建,结果与专家确定平面的角度误差仅为0.73°±0.50°。此外,深度学习还被应用于颈椎成熟度分期判断骨龄[41],以及辅助治疗决策,如决定牙齿是否拔除等[42]。
(5)牙周炎:牙周炎的自动诊断研究主要基于根尖片及曲面体层X线片。Lee等[43]开发了基于深度学习的根尖片牙周炎自动诊断系统,前磨牙牙周炎诊断准确度为0.81,磨牙为0.77。Kim等[44]开发了DeNTNet模型用于曲面体层X线片自动检测牙槽骨吸收,经过经验丰富的口腔医师标注的12 179张曲面体层X线片的训练,DeNTNet模型在测试集中F1分数可达0.75,高于口腔医师(0.69)。Chang等[45]将深度学习与传统计算机辅助诊断方法结合,用于曲面体层X线片自动检测牙周炎及其分期,结果与影像医师的Pearson相关系数达0.73。
(6)口腔种植:深度学习在口腔种植中的应用主要包括牙槽骨量分析、种植体系统识别及种植术后并发症诊断[46, 47, 48, 49, 50]。由于市场上存在许多品牌及型号的种植体,如何准确分辨这些种植体具有重要意义。口腔种植深度学习应用的一个热点即通过根尖片和曲面体层X线片识别不同种植体系统[47, 48],已有研究显示深度学习系统对种植体分类的准确性超过参与研究的多数医师[51]。在种植术后并发症的诊断方面,有学者探索深度学习用于种植体周炎及种植体折断的检测[49, 50],准确度较高。
(7)其他疾病:如颞下颌关节疾病、颌骨骨折、上颌窦炎症、骨质疏松、舍格伦综合征等影像学诊断中亦有深度学习的相关报道[52, 53, 54, 55]。颞下颌关节疾病方面,深度学习被用于曲面体层X线片及锥形束CT骨关节病的诊断[56, 57]。较多口腔医师对颞下颌关节盘移位的MRI诊断并不熟悉,已有学者报道基于CNN的MRI颞下颌关节盘移位自动诊断模型,准确度可达0.77[58]。
3.法医学牙齿鉴定:深度学习也可用于法医学牙齿鉴定,包括年龄推断、个人识别等。年龄推断方面,采用CNN评估曲面体层X线片智齿发育阶段可判断年龄[59]。Mu和Li[60]收集了3 000张曲面体层X线片用于训练4种迁移学习模型以判断年龄,其中表现最优的EfficientNet-B5模型的平均绝对误差为2.83,均方根误差为4.59。Zheng等[61]采用在锥形束CT中自动分割并计算第一磨牙髓腔体积的方法估计年龄,结果与真实年龄的相关系数r为0.74。Patil等[62]利用曲面体层X线片的下颌骨形态参数确定性别,准确度达0.75。
1.研究结果的稳健性和泛化性缺乏验证:尽管目前口腔颌面影像领域报道的深度学习模型具有良好的性能,但其可推广性尚未得到充分验证。这些研究主要聚焦于人工智能模型的研发,而关于模型稳健性、泛化性的研究较少。目前研究中用于开发人工智能模型的数据集常较小,且多为单中心研究,使用同一机构的图像通过交叉验证进行测试,很可能导致模型的过拟合。已有学者报道深度学习模型在其他机构的图像上测试时性能变差[63, 64]。
2.口腔颌面影像领域深度学习研究的质量有待提升:系统综述显示,口腔颌面影像领域人工智能研究存在较高的风险偏倚[17]。部分研究的数据集来源和数据特征缺乏详细描述,数据标注方法、金标准制订过程不够清晰。此外,部分人工智能模型的选择、训练和参数调整、验证策略不清晰。
3.深度学习的不透明性及脆弱性:深度学习模型自动提取复杂的数据特征,并从原始数据中优化加权参数,学者无法推断其决策过程,被认为是黑箱模型。算法的不透明性可影响深度学习模型在临床的使用与人工智能设备的监管。同时,深度学习系统具有脆弱性,易受到对抗性攻击,即通过向正常数据添加精细设计的微小改变即可导致模型作出明显错误的判断[65],给临床应用带来安全隐患。
4.研究任务较单一:目前口腔颌面影像领域人工智能研究多基于单一影像手段对单一类别的疾病进行诊断,而在临床实际诊疗中,医师通常需要结合不同影像学手段以及临床信息进行综合诊断。此外,多数研究针对的是临床常见病,而针对少见病、罕见病的研究较少。多数口腔医师对罕见病的诊断经验较少,针对此类疾病开发人工智能诊断模型也具有重要意义。
5.伦理与法规:医学影像领域人工智能产品的开发及临床验证需要大量的医疗数据,患者的隐私权需要得到保障。此外,基于深度学习的人工智能产品具有黑箱属性,且更新迭代快,对人工智能产品的评价和监管具有特殊性,需要相关法律法规的完善。
1.建设大样本、高质量、标准化数据库:建设大样本、高质量、标准化数据库对人工智能产品的研发及临床验证均有重要的意义。深度学习模型通过不同中心、不同时点的大样本数据训练,才能学习到病变的关键图像特征,避免过拟合,才具有可推广性。我国临床患者众多,口腔颌面影像数据量大,但缺乏完整的结构化数据。数据集的收集及人工标注费工费时、缺乏统一标准。提升数据标注规范性与准确性、提高标注效率,是目前研究亟待解决的问题。建议出台相应的数据标注指南,规范数据标注方法、对标注人员进行审核、培训,对标注流程进行质量控制,增加金标准的可靠性。同时加强不同区域、各级医院的合作,建设大样本、多样化、标准化、精标注数据库。
2.进一步提升口腔领域人工智能研究及报道的质量:针对目前部分医学影像人工智能研究的不足,部分专家及学者提出医学影像人工智能研究指南[66, 67],从研究设计、实施到结果分析、论文报道撰写方面为研究人员提供指导,以提高研究的严谨性与规范性。未来口腔领域人工智能研究应遵循标准化的报道格式,以更严格的科学标准进行研究设计,减少偏倚的风险,便于研究成果的评估和推广。
3.人工智能新技术的开发:针对训练样本不足、数据标注困难的问题,除优化数据集外,还可通过新技术的开发弥补数据不足的问题,比如迁移学习、图像增广技术、半监督学习算法、小样本学习、元学习、联邦学习等。联邦学习旨在通过多中心协作共同训练模型以解决数据不足和隐私问题,各中心之间仅共享模型参数,而不交换患者数据,这既可保护数据的隐私性,又可提高模型的泛化性能[68]。针对深度学习模型缺乏可解释性的问题,可应用梯度加权类激活映射生成热图,显示CNN分类时关注的焦点区域,提升深度学习模型的可解释性[69]。对于深度学习模型的脆弱性,可进行对抗性训练或采用复杂性较低的模型以增加模型的稳健性[70]。
4.从单一疾病的人工智能诊断模型向综合医疗辅助系统发展:目前的口腔颌面影像人工智能诊断研究多基于单一影像手段对单一类别疾病进行检测及诊断,与临床实际工作场景不同。在目前的工作基础上,需要对不同检查手段、不同疾病的诊断模型进行整合,包括:①同一影像检查中多种疾病的诊断;②不同影像手段用于同一种疾病的诊断;③影像资料与临床信息、实验室检查等信息结合的综合诊断。此外,人工智能系统应与医院影像归档和通信系统、信息系统进行整合,实现从以单病种为中心到以患者为中心的人工智能辅助系统。展望口腔医学领域人工智能的未来,不仅用于辅助疾病诊断,人工智能或将贯穿口腔医疗服务全过程,包括疾病预防、早期筛查、辅助治疗、判断预后、复查和监测、优化医疗工作流程等,以全面提升口腔医疗服务水平。