A cycle generative adversarial network for generating synthetic contrast-enhanced computed tomographic images from non-contrast images in the internal jugular lymph node-bearing area.

The objectives of this study were to create a mutual conversion system between contrast-enhanced computed tomography (CECT) and non-CECT images using a cycle generative adversarial network (cycleGAN) for the internal jugular region. Image patches were cropped from CT images in 25 patients who underwent both CECT and non-CECT imaging. Using a cycleGAN, synthetic CECT and non-CECT images were generated from original non-CECT and CECT images, respectively. The peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) were calculated. Visual Turing tests were used to determine whether oral and maxillofacial radiologists could tell the difference between synthetic versus original images, and receiver operating characteristic (ROC) analyses were used to assess the radiologists' performances in discriminating lymph nodes from blood vessels. The PSNR of non-CECT images was higher than that of CECT images, while the SSIM was higher in CECT images. The Visual Turing test showed a higher perceptual quality in CECT images. The area under the ROC curve showed almost perfect performances in synthetic as well as original CECT images. In conclusion, synthetic CECT images created by cycleGAN appeared to have the potential to provide effective information in patients who could not receive contrast enhancement.

Deep learning system for distinguishing between nasopalatine duct cysts and radicular cysts arising in the midline region of the anterior maxilla on panoramic radiographs.

Purpose: The aims of this study were to create a deep learning model to distinguish between nasopalatine duct cysts (NDCs), radicular cysts, and no-lesions (normal) in the midline region of the anterior maxilla on panoramic radiographs and to compare its performance with that of dental residents. Materials and Methods: One hundred patients with a confirmed diagnosis of NDC (53 men, 47 women; average age, 44.6±16.5 years), 100 with radicular cysts (49 men, 51 women; average age, 47.5±16.4 years), and 100 with normal groups (56 men, 44 women; average age, 34.4±14.6 years) were enrolled in this study. Cases were randomly assigned to the training datasets (80%) and the test dataset (20%). Then, 20% of the training data were randomly assigned as validation data. A learning model was created using a customized DetectNet built in Digits version 5.0 (NVIDIA, Santa Clara, USA). The performance of the deep learning system was assessed and compared with that of two dental residents. Results: The performance of the deep learning system was superior to that of the dental residents except for the recall of radicular cysts. The areas under the curve (AUCs) for NDCs and radicular cysts in the deep learning system were significantly higher than those of the dental residents. The results for the dental residents revealed a significant difference in AUC between NDCs and normal groups. Conclusion: This study showed superior performance in detecting NDCs and radicular cysts and in distinguishing between these lesions and normal groups.

Differences in the panoramic appearance of cleft alveolus patients with or without a cleft palate.

Purpose: The purpose of this study was to clarify the panoramic image differences of cleft alveolus patients with or without a cleft palate, with emphases on the visibility of the line formed by the junction between the nasal septum and nasal floor (the upper line) and the appearances of the maxillary lateral incisor. Materials and Methods: Panoramic radiographs of 238 patients with cleft alveolus were analyzed for the visibility of the upper line, including clear, obscure or invisible, and the appearances of the maxillary lateral incisor, regarding congenital absence, incomplete growth, delayed eruption and medial inclination. Differences in the distribution ratio of these visibility and appearances were verified between the patients with and without a cleft palate using the chi-square test. Results: There was a significant difference in the visibility distribution of the upper line between the patients with and without a cleft palate (p<0.05). In most of the patients with a cleft palate, the upper line was not observed. In the unilateral cleft alveolus patients, the medial inclination of the maxillary lateral incisor was more frequently observed in patients with a cleft palate than in patients without a cleft palate. Conclusion: Two differences were identified in panoramic appearances. The first was the disappearance (invisible appearance) of the upper line in patients with a cleft palate, and the second was a change in the medial inclination on the affected side maxillary lateral incisor in unilateral cleft alveolus patients with a cleft palate.

External Validation of the Effect of the Combined Use of Object Detection for the Classification of the C-Shaped Canal Configuration of the Mandibular Second Molar in Panoramic Radiographs: A Multicenter Study.

INTRODUCTION: The purposes of this study were to evaluate the effect of the combined use of object detection for the classification of the C-shaped canal anatomy of the mandibular second molar in panoramic radiographs and to perform an external validation on a multicenter dataset. METHODS: The panoramic radiographs of 805 patients were collected from 4 institutes across two countries. The CBCT data of the same patients were used as "Ground-truth". Five datasets were generated: one for training and validation, and 4 as external validation datasets. Workflow 1 used manual cropping to prepare the image patches of mandibular second molars, and then classification was performed using EfficientNet. Workflow 2 used two combined methods with a preceding object detection (YOLOv7) performed for automated image patch formation, followed by classification using EfficientNet. Workflow 3 directly classified the root canal anatomy from the panoramic radiographs using the YOLOv7 prediction outcomes. The classification performance of the 3 workflows was evaluated and compared across 4 external validation datasets. RESULTS: For Workflows 1, 2, and 3, the area under the receiver operating characteristic curve (AUC) values were 0.863, 0.861, and 0.876, respectively, for the AGU dataset; 0.935, 0.945, and 0.863, respectively, for the ASU dataset; 0.854, 0.857, and 0.849, respectively, for the ODU dataset; and 0.821, 0.797, and 0.831, respectively, for the ODU low-resolution dataset. No significant differences existed between the AUC values of Workflows 1, 2, and 3 across the 4 datasets. CONCLUSIONS: The deep learning systems of the 3 workflows achieved significant accuracy in predicting the C-shaped canal in mandibular second molars across all test datasets.

Metastatic Lymph Node Detection on Ultrasound Images Using YOLOv7 in Patients with Head and Neck Squamous Cell Carcinoma.

Ultrasonography is the preferred modality for detailed evaluation of enlarged lymph nodes (LNs) identified on computed tomography and/or magnetic resonance imaging, owing to its high spatial resolution. However, the diagnostic performance of ultrasonography depends on the examiner's expertise. To support the ultrasonographic diagnosis, we developed YOLOv7-based deep learning models for metastatic LN detection on ultrasonography and compared their detection performance with that of highly experienced radiologists and less experienced residents. We enrolled 462 B- and D-mode ultrasound images of 261 metastatic and 279 non-metastatic histopathologically confirmed LNs from 126 patients with head and neck squamous cell carcinoma. The YOLOv7-based B- and D-mode models were optimized using B- and D-mode training and validation images and their detection performance for metastatic LNs was evaluated using B- and D-mode testing images, respectively. The D-mode model's performance was comparable to that of radiologists and superior to that of residents' reading of D-mode images, whereas the B-mode model's performance was higher than that of residents but lower than that of radiologists on B-mode images. Thus, YOLOv7-based B- and D-mode models can assist less experienced residents in ultrasonographic diagnoses. The D-mode model could raise the diagnostic performance of residents to the same level as experienced radiologists.

Effect of deep transfer learning with a different kind of lesion on classification performance of pre-trained model: Verification with radiolucent lesions on panoramic radiographs.

Purpose: The aim of this study was to clarify the influence of training with a different kind of lesion on the performance of a target model. Materials and Methods: A total of 310 patients (211 men, 99 women; average age, 47.9±16.1 years) were selected and their panoramic images were used in this study. We created a source model using panoramic radiographs including mandibular radiolucent cyst-like lesions (radicular cyst, dentigerous cyst, odontogenic keratocyst, and ameloblastoma). The model was simulatively transferred and trained on images of Stafne's bone cavity. A learning model was created using a customized DetectNet built in the Digits version 5.0 (NVIDIA, Santa Clara, CA). Two machines (Machines A and B) with identical specifications were used to simulate transfer learning. A source model was created from the data consisting of ameloblastoma, odontogenic keratocyst, dentigerous cyst, and radicular cyst in Machine A. Thereafter, it was transferred to Machine B and trained on additional data of Stafne's bone cavity to create target models. To investigate the effect of the number of cases, we created several target models with different numbers of Stafne's bone cavity cases. Results: When the Stafne's bone cavity data were added to the training, both the detection and classification performances for this pathology improved. Even for lesions other than Stafne's bone cavity, the detection sensitivities tended to increase with the increase in the number of Stafne's bone cavities. Conclusion: This study showed that using different lesions for transfer learning improves the performance of the model.

Alterations of posterior pharyngeal wall movement during swallowing in postoperative tongue cancer patients: assessment with a videofluoroscopic swallowing study.

This study aimed to determine the association between the progressive contraction of the posterior pharyngeal wall and dysphagia in postoperative patients with tongue cancer. A videofluoroscopic swallowing study (VFSS) was performed in 34 patients after tongue cancer surgery. Images were analyzed using a two-dimensional video measurement software. Cases in which the processes on the posterior pharyngeal wall moved downward from the 2nd to 4th vertebral regions were defined as "normal type", other cases were defined as "abnormal type". Twenty-four patients showed normal movement of the posterior pharyngeal wall, whereas 10 patients showed the abnormal type. The results showed that there was a significant difference in dysphagia scores between the postoperative swallowing type and swallowing dysfunction score. This implies that dysphagia is related to the movement of the posterior pharyngeal wall after tongue cancer surgery. Furthermore, the extent of resection and stage were significantly different between the normal and abnormal groups in the posterior pharyngeal wall movement. There was also a significant difference between the two groups in terms of the following: whether the tongue base was included in the excision range (p < 0.01), whether neck dissection was performed (p < 0.01), or whether reconstruction was not performed (p < 0.01). VFSS results showed that posterior pharyngeal wall movement was altered after surgery in patients with tongue cancer who had severe dysphagia.

Detection of unilateral and bilateral cleft alveolus on panoramic radiographs using a deep-learning system.

OBJECTIVES: The purpose of this study was to evaluate the difference in performance of deep-learning (DL) models with respect to the image classes and amount of training data to create an effective DL model for detecting both unilateral cleft alveoli (UCAs) and bilateral cleft alveoli (BCAs) on panoramic radiographs. METHODS: Model U was created using UCA and normal images, and Model B was created using BCA and normal images. Models C1 and C2 were created using the combined data of UCA, BCA, and normal images. The same number of CAs was used for training Models U, B, and C1, whereas Model C2 was created with a larger amount of data. The performance of all four models was evaluated with the same test data and compared with those of two human observers. RESULTS: The recall values were 0.60, 0.73, 0.80, and 0.88 for Models A, B, C1, and C2, respectively. The results of Model C2 were highest in precision and F-measure (0.98 and 0.92) and almost the same as those of human observers. Significant differences were found in the ratios of detected to undetected CAs of Models U and C1 (p = 0.01), Models U and C2 (p < 0.001), and Models B and C2 (p = 0.036). CONCLUSIONS: The DL models trained using both UCA and BCA data (Models C1 and C2) achieved high detection performance. Moreover, the performance of a DL model may depend on the amount of training data.

Deep-learning systems for diagnosing cleft palate on panoramic radiographs in patients with cleft alveolus.

OBJECTIVES: The aim of the present study was to create effective deep learning-based models for diagnosing the presence or absence of cleft palate (CP) in patients with unilateral or bilateral cleft alveolus (CA) on panoramic radiographs. METHODS: The panoramic images of 491 patients who had unilateral or bilateral cleft alveolus were used to create two models. Model A, which detects the upper incisor area on panoramic radiographs and classifies the areas into the presence or absence of CP, was created using both object detection and classification functions of DetectNet. Using the same data for developing Model A, Model B, which directly classifies the presence or absence of CP on panoramic radiographs, was created using classification function of VGG-16. The performances of both models were evaluated with the same test data and compared with those of two radiologists. RESULTS: The recall, precision, and F-measure were all 1.00 in Model A. The area under the receiver operating characteristic curve (AUC) values were 0.95, 0.93, 0.70, and 0.63 for Model A, Model B, and the radiologists, respectively. The AUCs of the models were significantly higher than those of the radiologists. CONCLUSIONS: The deep learning-based models developed in the present study have potential for use in supporting observer interpretations of the presence of cleft palate on panoramic radiographs.

Transfer learning in diagnosis of maxillary sinusitis using panoramic radiography and conventional radiography.

OBJECTIVES: To clarify the performance of transfer learning with a small number of Waters' images at institution B in diagnosing maxillary sinusitis, based on a source model trained with a large number of panoramic radiographs at institution A. METHODS: The source model was created by a 200-epoch training process with 800 training and 60 validation datasets of panoramic radiographs at institution A using VGG-16. One hundred and eighty Waters' and 180 panoramic image patches with or without maxillary sinusitis at institution B were enrolled in this study, and were arbitrarily assigned to 120 training, 20 validation, and 40 test datasets, respectively. Transfer learning of 200 epochs was performed using the training and validation datasets of Waters' images based on the source model, and the target model was obtained. The test Waters' images were applied to the source and target models, and the performance of each model was evaluated. Transfer learning with panoramic radiographs and evaluation by two radiologists were undertaken and compared. The evaluation was based on the area of receiver-operating characteristic curves (AUC). RESULTS: When using Waters' images as the test dataset, the AUCs of the source model, target model, and radiologists were 0.780, 0.830, and 0.806, respectively. There were no significant differences between these models and the radiologists, whereas the target model performed better than the source model. For panoramic radiographs, AUCs were 0.863, 0.863, and 0.808, respectively, with no significant differences. CONCLUSIONS: This study performed transfer learning using a small number of Waters' images, based on a source model created solely from panoramic radiographs, resulting in a performance improvement to 0.830 in diagnosing maxillary sinusitis, which was equivalent to that of radiologists. Transfer learning is considered a useful method to improve diagnostic performance.

A preliminary deep learning study on automatic segmentation of contrast-enhanced bolus in videofluorography of swallowing.

Although videofluorography (VFG) is an effective tool for evaluating swallowing functions, its accurate evaluation requires considerable time and effort. This study aimed to create a deep learning model for automated bolus segmentation on VFG images of patients with healthy swallowing and dysphagia using the artificial intelligence deep learning segmentation method, and to assess the performance of the method. VFG images of 72 swallowing of 12 patients were continuously converted into 15 static images per second. In total, 3910 images were arbitrarily assigned to the training, validation, test 1, and test 2 datasets. In the training and validation datasets, images of colored bolus areas were prepared, along with original images. Using a U-Net neural network, a trained model was created after 500 epochs of training. The test datasets were applied to the trained model, and the performances of automatic segmentation (Jaccard index, Sørensen-Dice coefficient, and sensitivity) were calculated. All performance values for the segmentation of the test 1 and 2 datasets were high, exceeding 0.9. Using an artificial intelligence deep learning segmentation method, we automatically segmented the bolus areas on VFG images; our method exhibited high performance. This model also allowed assessment of aspiration and laryngeal invasion.

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.

OBJECTIVE: The aim of this study was to create and assess a deep learning model using segmentation and transfer learning methods to visualize the proximity of the mandibular canal to an impacted third molar on panoramic radiographs. STUDY DESIGN: The panoramic radiographs containing the mandibular canal and impacted third molar were collected from 2 hospitals (Hospitals A and B). A total of 3200 areas were used for creating and evaluating learning models. A source model was created using the data from Hospital A, simulatively transferred to Hospital B, and trained using various amounts of data from Hospital B to create target models. The same data were then applied to the target models to calculate the Dice coefficient, Jaccard index, and sensitivity. RESULTS: The performance of target models trained using 200 or more data sets was equivalent to that of the source model tested using data obtained from the same hospital (Hospital A). CONCLUSIONS: Sufficiently qualified models could delineate the mandibular canal in relation to an impacted third molar on panoramic radiographs using a segmentation technique. Transfer learning appears to be an effective method for creating such models using a relatively small number of data sets.

Effect of a lead block on alveolar bone protection in image-guided high-dose-rate interstitial brachytherapy for tongue cancer: using model-based dose calculation algorithms to correct for inhomogeneity.

PURPOSE: The purpose of this study was to evaluate the effect of a lead block for alveolar bone protection in image-guided high-dose-rate interstitial brachytherapy for tongue cancer. MATERIAL AND METHODS: We treated 6 patients and delivered 5,400 cGy in 9 fractions using a lead block. Effects of lead block (median thickness, 4 mm) on dose attenuation by distance were visually examined using TG-43 formalism-based dose distribution curves to determine whether or not the area with the highest dose is located in the alveolar bone, where there is a high-risk of infection. Dose re-calculations were performed using TG-186 formalism with advanced collapsed cone engine (ACE) for inhomogeneity correction set to cortical bone density for the whole mandible and alveolar bone, water density for clinical target volume (CTV), air density for outside body and lead density, and silastic density for lead block and its' silicon replica, respectively. RESULTS: The highest dose was detected outside the alveolar bone in five of the six cases. For dose-volume histogram analysis, median minimum doses delivered per fraction to the 0.1 cm3 of alveolar bone (D0.1cm3 TG-43, ACE-silicon, and ACE-lead) were 344.3 (range, 262.9-427.4) cGy, 336.6 (253.3-425.0) cGy, and 169.7 (114.9-233.3) cGy, respectively. D0.1cm3 ACE-lead was significantly lower than other parameters. No significant difference was observed between CTV-related parameters. CONCLUSIONS: The results suggested that using a lead block for alveolar bone protection with a thickness of about 4 mm, can shift the highest dose area to non-alveolar regions. In addition, it reduced D0.1cm3 of alveolar bone to about half, without affecting tumor dose.

Segmentation of metastatic cervical lymph nodes from CT images of oral cancers using deep-learning technology.

OBJECTIVE: The purpose of this study was to establish a deep-learning model for segmenting the cervical lymph nodes of oral cancer patients and diagnosing metastatic or non-metastatic lymph nodes from contrast-enhanced computed tomography (CT) images. METHODS: CT images of 158 metastatic and 514 non-metastatic lymph nodes were prepared. CT images were assigned to training, validation, and test datasets. The colored images with lymph nodes were prepared together with the original images for the training and validation datasets. Learning was performed for 200 epochs using the neural network U-net. Performance in segmenting lymph nodes and diagnosing metastasis were obtained. RESULTS: Performance in segmenting metastatic lymph nodes showed recall of 0.742, precision of 0.942, and F1 score of 0.831. The recall of metastatic lymph nodes at level II was 0.875, which was the highest value. The diagnostic performance of identifying metastasis showed an area under the curve (AUC) of 0.950, which was significantly higher than that of radiologists (0.896). CONCLUSIONS: A deep-learning model was created to automatically segment the cervical lymph nodes of oral squamous cell carcinomas. Segmentation performances should still be improved, but the segmented lymph nodes were more accurately diagnosed for metastases compared with evaluation by humans.

Orthodontic tooth movement-activated sensory neurons contribute to enhancing osteoclast activity and tooth movement through sympathetic nervous signalling.

OBJECTIVES: Orthodontic tooth movement (OTM) increases sympathetic and sensory neurological markers in periodontal tissue. However, the relationship between the sympathetic and sensory nervous systems during OTM remains unclear. Therefore, the present study investigated the relationship between the sympathetic and sensory nervous systems activated by OTM using pharmacological methods. MATERIALS AND METHODS: We compared the effects of sympathectomy and sensory nerve injury during OTM in C57BL6/J mice. Capsaicin (CAP) was used to induce sensory nerve injury. Sympathectomy was performed using 6-hydroxydopamine. To investigate the effects of a ß-agonist on sensory nerve injury, isoproterenol (ISO) was administered to CAP-treated mice. Furthermore, to examine the role of the central nervous system in OTM, the ventromedial hypothalamic nucleus (VMH) was ablated using gold thioglucose. RESULTS: Sensory nerve injury and sympathectomy both suppressed OTM and decreased the percent of the alveolar socket covered with osteoclasts (Oc.S/AS) in periodontal tissue. Sensory nerve injury inhibited increases in OTM-induced calcitonin gene-related peptide (CGRP) immunoreactivity (IR), a marker of sensory neurons, and tyrosine hydroxylase (TH) IR, a marker of sympathetic neurons, in periodontal tissue. Although sympathectomy did not decrease the number of CGRP-IR neurons in periodontal tissue, OTM-induced increases in the number of TH-IR neurons were suppressed. The ISO treatment restored sensory nerve injury-inhibited tooth movement and Oc.S/AS. Furthermore, the ablation of VMH, the centre of the sympathetic nervous system, suppressed OTM-induced increases in tooth movement and Oc.S/AS. CONCLUSIONS: The present results suggest that OTM-activated sensory neurons contribute to enhancements in osteoclast activity and tooth movement through sympathetic nervous signalling.

Efficacy of a deep leaning model created with the transfer learning method in detecting sialoliths of the submandibular gland on panoramic radiography.

OBJECTIVE: This study aimed to compare the performance of 3 deep learning models, including a model constructed with the transfer learning method, in detecting submandibular gland sialoliths on panoramic radiographs. STUDY DESIGN: We used data from 2 institutions (A and B) to create the models for use in institution B. In total, 224 panoramic radiographs with sialoliths were used. Model 1 was created using data from institution A only, model 2 was created using combined data from institutions A and B, and model 3 was created using the transfer learning method by having model 1 transferred and trained in various learning epochs using data from institution B. These models were tested and compared in their detection performance using testing data sets from institution B. RESULTS: Model 2 and model 3 with 300 epochs performed equally well and yielded the highest detection rates (recall: sensitivity of 85%, precision: positive predictive value of 100%, and F measure of 91.9%) for sialoliths on panoramic radiographs. CONCLUSION: The results of this study suggest that use of the transfer learning method with an appropriate number of epochs may be an alternative to sharing patient personal data among institutions.

Automatic segmentation of the temporomandibular joint disc on magnetic resonance images using a deep learning technique.

OBJECTIVES: The aims of the present study were to construct a deep learning model for automatic segmentation of the temporomandibular joint (TMJ) disc on magnetic resonance (MR) images, and to evaluate the performances using the internal and external test data. METHODS: In total, 1200 MR images of closed and open mouth positions in patients with temporomandibular disorder (TMD) were collected from two hospitals (Hospitals A and B). The training and validation data comprised 1000 images from Hospital A, which were used to create a segmentation model. The performance was evaluated using 200 images from Hospital A (internal validity test) and 200 images from Hospital B (external validity test). RESULTS: Although the analysis of performance determined with data from Hospital B showed low recall (sensitivity), compared with the performance determined with data from Hospital A, both performances were above 80%. Precision (positive predictive value) was lower when test data from Hospital A were used for the position of anterior disc displacement. According to the intra-articular TMD classification, the proportions of accurately assigned TMJs were higher when using images from Hospital A than when using images from Hospital B. CONCLUSION: The segmentation deep learning model created in this study may be useful for identifying disc positions on MR images.

Performance of deep learning technology for evaluation of positioning quality in periapical radiography of the maxillary canine.

OBJECTIVES: The aim of the present study was to create and test an automatic system for assessing the technical quality of positioning in periapical radiography of the maxillary canines using deep learning classification and segmentation techniques. METHODS: We created and tested two deep learning systems using 500 periapical radiographs (250 each of good- and bad-quality images). We assigned 350, 70, and 80 images as the training, validation, and test datasets, respectively. The learning model of system 1 was created with only the classification process, whereas system 2 consisted of both the segmentation and classification models. In each model, 500 epochs of training were performed using AlexNet and U-net for classification and segmentation, respectively. The segmentation results were evaluated by the intersection over union method, with values of 0.6 or more considered as success. The classification results were compared between the two systems. RESULTS: The segmentation performance of system 2 was recall, precision, and F measure of 0.937, 0.961, and 0.949, respectively. System 2 showed better classification performance values than those obtained by system 1. The area under the receiver operating characteristic curve values differed significantly between system 1 (0.649) and system 2 (0.927). CONCLUSIONS: The deep learning systems we created appeared to have potential benefits in evaluation of the technical positioning quality of periapical radiographs through the use of segmentation and classification functions.

Preliminary Study on the Diagnostic Performance of a Deep Learning System for Submandibular Gland Inflammation Using Ultrasonography Images.

This study was performed to evaluate the diagnostic performance of deep learning systems using ultrasonography (USG) images of the submandibular glands (SMGs) in three different conditions: obstructive sialoadenitis, Sjögren's syndrome (SjS), and normal glands. Fifty USG images with a confirmed diagnosis of obstructive sialoadenitis, 50 USG images with a confirmed diagnosis of SjS, and 50 USG images with no SMG abnormalities were included in the study. The training group comprised 40 obstructive sialoadenitis images, 40 SjS images, and 40 control images, and the test group comprised 10 obstructive sialoadenitis images, 10 SjS images, and 10 control images for deep learning analysis. The performance of the deep learning system was calculated and compared between two experienced radiologists. The sensitivity of the deep learning system in the obstructive sialoadenitis group, SjS group, and control group was 55.0%, 83.0%, and 73.0%, respectively, and the total accuracy was 70.3%. The sensitivity of the two radiologists was 64.0%, 72.0%, and 86.0%, respectively, and the total accuracy was 74.0%. This study revealed that the deep learning system was more sensitive than experienced radiologists in diagnosing SjS in USG images of two case groups and a group of healthy subjects in inflammation of SMGs.