Opinion: The letter to Editor regarding "Automatic segmentation of the maxillary sinus on cone beam computed tomographic images with U-Net deep learning model"

Jiayi Chen^1,2*

• ¹Jiayi Chen, Suzhou, China
• ²Suzhou Wujiang District Hospital of Traditional Chinese Medicine, suzhou, China

With the rapid development of technology in modern society, AI (Artificial Intelligence), a product of the new generation of technological revolution, gradually plays a vital role on various fields. The term "AI" refers to that a machine is able to perform tasks like humans. ML (Machine Learning), as a subdomain of AI, is an algorithm that it can predict the results from unseen data by learning intrinsic statistical pattern in data. DL (Deep Learning), as one of ML algorithms, uses multi-layer calculations to learn and infer on data, especially images [1]. Over the last few years, DL has put deep effects on several different scientific disciplines. In addition, DL has been applied for the state-of-the-art tasks like the autonomous driving. For example, Tesla can perform autonomous driving on the expressway without supervision. DL also demonstrates vast potential in biomedical domain. According to PubMed searching results, biomedical studies generally focus on special medical scenarios, including the analysis of medical images such as CT and pathological images [2]. However, even though PubMed is the largest search engine in medical domain at present, it cannot cover all medicine-related publications. For instance, medical publications are also collected by computer science researches, most often conferences [3][4]. The prestigious annual conference like MICCAI (Medical Image Computing and Computer Assisted Intervention) have very influential effects on biomedical domain. Compare to medicine, which has a history of a millennia-old tradition, computer science seems to be a young discipline. Nevertheless, the combination of computer science and medicine will bring revolutionary changes in medical domain.Generally, paranasal sinuses consist of four parts, including the maxillary, ethmoid, frontal and sphenoid. They are full of air, lined with mucosa, and communicating with the nasal cavity. Maxillary sinus was first reported by Leonardo da Vinci in 1489 and documented by the English anatomist Nathaniel Highmore in 1651. Hence, it also can be called as "Highmore" sinus [5].Maxillary sinus is a pyramid-shaped cavity located on the maxilla and alleviates the weight of maxilla. It is no longer acceptable to consider maxillary sinus as two parts, containing an inferior part supporting the dentoalveolar process that is the focus of dental practitioner and a superior part related to sinus ostium which is the focus of ENT specialists. Therefore, how to recognize the boundary of maxillary sinus on CT/CBCT is vital to OMFS and ENT specialists. For example, maxillary sinus floor elevation is a classic technology in dental implantology, requiring dental practitioners to master the anatomy of maxillary sinus floor. They should judge whether its floor is suitable to perform this surgery on CBCT images. Maxillary sinus puncture procedure is a classic surgery to alleviate the inflammation of maxillary sinus for ENT specialists. It is important for them to evaluate the degree of inflammation of maxillary sinus on CT images. However, how to read and recognized the anatomy and diseases of maxillary sinus on CT images is a challenge to young medical and dental practitioners.Recently, DL has been applied to the diagnosis and identification of maxillary sinus on CT/CBCT [6][7]. This state-of-the-art technology demonstrates the potential of diagnosis of diseases for clinical practitioners. In general, the segmentation of regions of interest (ROIs) is an essential step of DL. However, manual delineation of ROI by physicians on CBCT/CT images is time-consuming and subjective, with the potential bias between observers [8]. Deep learning-based auto-segmentation depends on the availability of annotated datasets. It obeys to the general criteria: the more data there are, the more accuracy the results [9]. However, high-quality medical images are scare as it requires operator to master special medical knowledge and take ethical principles into consideration. Therefore, it is vital to estimate the required training sample size before DL model is applied on medical diagnosis. Insufficient sample size can lead to overfitting of training sets and underfitting of target tasks. If the sample size can be estimated by the setting performance goals, it might help researchers better design. The air chamber in maxillary sinus communicates with the nasal cavity via an ostiomeatal complex, which is made up of one or more ostia opening into the semilunar hiatus of the middle nasal meatus [10]. Generally, it is difficult to distinguish the maxillary sinus and sinus ostium in medical images. Furthermore, the mucosa was thickened by the inflammation of maxillary sinus, which can lead to stenosis of the sinus opening. It is still a challenge to clinical physicians to annotate the boundary of maxillary sinus. Meanwhile, there are anatomic variation of maxillary sinus, including maxillary septa, maxillary sinus dysplasia, and maxillary antroliths. These factors can affect the accuracy of annotation of the maxillary sinus, resulting the bias of segmentation using DL model. I wholeheartedly commend Busra et al. academic contribution entitled "Automatic segmentation of the maxillary sinus on cone beam computed tomographic images with U-Net deep learning model" [11]. This novelty application is set to improve the development of otolaryngology and dentistry. However, I have some confusions on this study from the perspective of a dentist.The maxillary sinus is not a fully cavity located in maxilla and maxillary sinus septa and maxillary antrolith are usually discovered in Cone Beam Computerized Tomography based on our experience. Septa and antrolith are pivotal factors for maxillary sinus floor elevation in implant dentistry, which can lead to Schneider membrane perforation during surgery procedure. It is no doubt that automatic segmentation of maxillary on CBCT images based on deep learning algorithm will reduce the rate of membrane perforation. However, in this study author did not mention these factors in inclusion criteria and exclusion criteria. Whether U-net frame can predict the different maxillary sinus with high mIoU/accuracy or not deserve further research in the future.Secondly, a prior study revealed that the train sample size has a significant influence on the performance of deep learning model for head-neck organ segmentation [12]. The optimal sample size for deep learning model still remains controversy. Unet frame was first proposed in 2015 and famous for its low demand for sample size in medical image segmentation. In this study, 100 axis CBCT images were analyzed and achieved 0.9275 (IoU value) and 0.9784 (F1 score) by U-net frame. Different maxillary sinus planes and less train sample size might be further investigated in the future.In terms of limitation, author mentioned that data from multi-centers could enhance the generalization ability of the model. I recommend CBCT images gathered from different nations and regions. A call for established public CBCT database might foster global research on this topic.In summary, this study is well-designed and novelty. The integration of deep learning into maxillary sinus diagnosis is promising, and assist otolaryngologists, oral and maxillofacial surgeon to diagnose with accuracy. Limitations of this model require our further investigation.