OBJECTIVE: To investigate a new noninvasive diagnostic model for nonalcoholic fatty liver disease (NAFLD) based on features of tongue images. METHODS: Healthy controls and volunteers confirmed to have NAFLD by liver ultrasound were recruited from China-Japan Friendship Hospital between September 2018 and May 2019, then the anthropometric indexes and sampled tongue images were measured. The tongue images were labeled by features, based on a brief protocol, without knowing any other clinical data, after a series of corrections and data cleaning. The algorithm was trained on images using labels and several anthropometric indexes for inputs, utilizing machine learning technology. Finally, a logistic regression algorithm and a decision tree model were constructed as 2 diagnostic models for NAFLD. RESULTS: A total of 720 subjects were enrolled in this study, including 432 patients with NAFLD and 288 healthy volunteers. Of them, 482 were randomly allocated into the training set and 238 into the validation set. The diagnostic model based on logistic regression exhibited excellent performance: in validation set, it achieved an accuracy of 86.98%, sensitivity of 91.43%, and specificity of 80.61%; with an area under the curve (AUC) of 0.93 [95% confidence interval (CI) 0.68-0.98]. The decision tree model achieved an accuracy of 81.09%, sensitivity of 91.43%, and specificity of 66.33%; with an AUC of 0.89 (95% CI 0.66-0.92) in validation set. CONCLUSIONS The features of tongue images were associated with NAFLD. Both the 2 diagnostic models, which would be convenient, noninvasive, lightweight, rapid, and inexpensive technical references for early screening, can accurately distinguish NAFLD and are worth further study.
Humans ; Non-alcoholic Fatty Liver Disease/diagnostic imaging* ; Ultrasonography ; Anthropometry ; Algorithms ; China

Humans ; Artificial Intelligence ; Algorithms ; Brain ; Head

In light of the increasing digitalization of dentistry, the automatic determination of three-dimensional (3D) craniomaxillofacial features has become a development trend. 3D craniomaxillofacial landmarks and symmetry reference plane determination algorithm based on point clouds has attracted a lot of attention, for point clouds are the basis for virtual surgery design and facial asymmetry analysis, which play a key role in craniomaxillofacial surgery and orthodontic treatment design. Based on the studies of our team and national and international literatures, this article presented the deep geometry learning algorithm to determine landmarks and symmetry reference plane based on 3D craniomaxillofacial point clouds. In order to provide reference for future clinical application, we describe the development and latest research in this field, and analyze and discuss the advantages and limitations of various methods.
Humans ; Imaging, Three-Dimensional/methods* ; Facial Asymmetry ; Algorithms

Objective: To construct a kind of neural network for eliminating the metal artifacts in CT images by training the generative adversarial networks (GAN) model, so as to provide reference for clinical practice. Methods: The CT data of patients treated in the Department of Radiology, West China Hospital of Stomatology, Sichuan University from January 2017 to June 2022 were collected. A total of 1 000 cases of artifact-free CT data and 620 cases of metal artifact CT data were obtained, including 5 types of metal restorative materials, namely, fillings, crowns, titanium plates and screws, orthodontic brackets and metal foreign bodies. Four hundred metal artifact CT data and 1 000 artifact-free CT data were utilized for simulation synthesis, and 1 000 pairs of simulated artifacts and metal images and simulated metal images (200 pairs of each type) were constructed. Under the condition that the data of the five metal artifacts were equal, the entire data set was randomly (computer random) divided into a training set (800 pairs) and a test set (200 pairs). The former was used to train the GAN model, and the latter was used to evaluate the performance of the GAN model. The test set was evaluated quantitatively and the quantitative indexes were root-mean-square error (RMSE) and structural similarity index measure (SSIM). The trained GAN model was employed to eliminate the metal artifacts from the CT data of the remaining 220 clinical cases of metal artifact CT data, and the elimination results were evaluated by two senior attending doctors using the modified LiKert scale. Results: The RMSE values for artifact elimination of fillings, crowns, titanium plates and screws, orthodontic brackets and metal foreign bodies in test set were 0.018±0.004, 0.023±0.007, 0.015±0.003, 0.019±0.004, 0.024±0.008, respectively (F=1.29, P=0.274). The SSIM values were 0.963±0.023, 0.961±0.023, 0.965±0.013, 0.958±0.022, 0.957±0.026, respectively (F=2.22, P=0.069). The intra-group correlation coefficient of 2 evaluators was 0.972. For 220 clinical cases, the overall score of the modified LiKert scale was (3.73±1.13), indicating a satisfactory performance. The scores of modified LiKert scale for fillings, crowns, titanium plates and screws, orthodontic brackets and metal foreign bodies were (3.68±1.13), (3.67±1.16), (3.97±1.03), (3.83±1.14), (3.33±1.12), respectively (F=1.44, P=0.145). Conclusions: The metal artifact reduction GAN model constructed in this study can effectively remove the interference of metal artifacts and improve the image quality.
Humans ; Tomography, X-Ray Computed/methods* ; Deep Learning ; Titanium ; Neural Networks, Computer ; Metals ; Image Processing, Computer-Assisted/methods* ; Algorithms

Objective: To explore an automatic landmarking method for anatomical landmarks in the three-dimensional (3D) data of the maxillary complex and preliminarily evaluate its reproducibility and accuracy. Methods: From June 2021 to December 2022, spiral CT data of 31 patients with relatively normal craniofacial morphology were selected from those who visited the Department of Oral and Maxillofacial Surgery, Peking University School and Hospital of Stomatology. The sample included 15 males and 16 females, with the age of (33.3±8.3) years. The maxillary complex was reconstructed in 3D using Mimics software, and the resulting 3D data of the maxillary complex was mesh-refined using Geomagic software. Two attending physicians and one associate chief physician manually landmarked the 31 maxillary complex datasets, determining 24 anatomical landmarks. The average values of the three expert landmarking results were used as the expert-defined landmarks. One case that conformed to the average 3D morphological characteristics of healthy individuals' craniofacial bones was selected as the template data, while the remaining 30 cases were used as target data. The open-source MeshMonk program (a non-rigid registration algorithm) was used to perform an initial alignment of the template and target data based on 4 landmarks (nasion, left and right zygomatic arch prominence, and anterior nasal spine). The template data was then deformed to the shape of the target data using a non-rigid registration algorithm, resulting in the deformed template data. Based on the unchanged index property of homonymous landmarks before and after deformation of the template data, the coordinates of each landmark in the deformed template data were automatically retrieved as the automatic landmarking coordinates of the homonymous landmarks in the target data, thus completing the automatic landmarking process. The automatic landmarking process for the 30 target data was repeated three times. The root-mean-square distance (RMSD) of the dense corresponding point pairs (approximately 25 000 pairs) between the deformed template data and the target data was calculated as the deformation error of the non-rigid registration algorithm, and the intra-class correlation coefficient (ICC) of the deformation error in the three repetitions was analyzed. The linear distances between the automatic landmarking results and the expert-defined landmarks for the 24 anatomical landmarks were calculated as the automatic landmarking errors, and the ICC values of the 3D coordinates in the three automatic landmarking repetitions were analyzed. Results: The average three-dimensional deviation (RMSD) between the deformed template data and the corresponding target data for the 30 cases was (0.70±0.09) mm, with an ICC value of 1.00 for the deformation error in the three repetitions of the non-rigid registration algorithm. The average automatic landmarking error for the 24 anatomical landmarks was (1.86±0.30) mm, with the smallest error at the anterior nasal spine (0.65±0.24) mm and the largest error at the left oribital (3.27±2.28) mm. The ICC values for the 3D coordinates in the three automatic landmarking repetitions were all 1.00. Conclusions: This study established an automatic landmarking method for three-dimensional data of the maxillary complex based on a non-rigid registration algorithm. The accuracy and repeatability of this method for landmarking normal maxillary complex 3D data were relatively good.
Male ; Female ; Humans ; Adult ; Imaging, Three-Dimensional/methods* ; Reproducibility of Results ; Algorithms ; Software ; Tomography, Spiral Computed ; Anatomic Landmarks/anatomy & histology*

Objective: To develop a multi-classification orthodontic image recognition system using the SqueezeNet deep learning model for automatic classification of orthodontic image data. Methods: A total of 35 000 clinical orthodontic images were collected in the Department of Orthodontics, Capital Medical University School of Stomatology, from October to November 2020 and June to July 2021. The images were from 490 orthodontic patients with a male-to-female ratio of 49∶51 and the age range of 4 to 45 years. After data cleaning based on inclusion and exclusion criteria, the final image dataset included 17 453 face images (frontal, smiling, 90° right, 90° left, 45° right, and 45° left), 8 026 intraoral images [frontal occlusion, right occlusion, left occlusion, upper occlusal view (original and flipped), lower occlusal view (original and flipped) and coverage of occlusal relationship], 4 115 X-ray images [lateral skull X-ray from the left side, lateral skull X-ray from the right side, frontal skull X-ray, cone-beam CT (CBCT), and wrist bone X-ray] and 684 other non-orthodontic images. A labeling team composed of orthodontic doctoral students, associate professors, and professors used image labeling tools to classify the orthodontic images into 20 categories, including 6 face image categories, 8 intraoral image categories, 5 X-ray image categories, and other images. The data for each label were randomly divided into training, validation, and testing sets in an 8∶1∶1 ratio using the random function in the Python programming language. The improved SqueezeNet deep learning model was used for training, and 13 000 natural images from the ImageNet open-source dataset were used as additional non-orthodontic images for algorithm optimization of anomaly data processing. A multi-classification orthodontic image recognition system based on deep learning models was constructed. The accuracy of the orthodontic image classification was evaluated using precision, recall, F1 score, and confusion matrix based on the prediction results of the test set. The reliability of the model's image classification judgment logic was verified using the gradient-weighted class activation mapping (Grad-CAM) method to generate heat maps. Results: After data cleaning and labeling, a total of 30 278 orthodontic images were included in the dataset. The test set classification results showed that the precision, recall, and F1 scores of most classification labels were 100%, with only 5 misclassified images out of 3 047, resulting in a system accuracy of 99.84%(3 042/3 047). The precision of anomaly data processing was 100% (10 500/10 500). The heat map showed that the judgment basis of the SqueezeNet deep learning model in the image classification process was basically consistent with that of humans. Conclusions: This study developed a multi-classification orthodontic image recognition system for automatic classification of 20 types of orthodontic images based on the improved SqueezeNet deep learning model. The system exhibitted good accuracy in orthodontic image classification.
Humans ; Male ; Female ; Child, Preschool ; Child ; Adolescent ; Young Adult ; Adult ; Middle Aged ; Deep Learning ; Reproducibility of Results ; Radiography ; Algorithms ; Cone-Beam Computed Tomography

OBJECTIVE: To propose a semi-supervised material quantitative intelligent imaging algorithm based on prior information perception learning (SLMD-Net) to improve the quality and precision of spectral CT imaging. METHODS: The algorithm includes a supervised and a self- supervised submodule. In the supervised submodule, the mapping relationship between low and high signal-to-noise ratio (SNR) data was constructed through mean square error loss function learning based on a small labeled dataset. In the self- supervised sub-module, an image recovery model was utilized to construct the loss function incorporating the prior information from a large unlabeled low SNR basic material image dataset, and the total variation (TV) model was used to to characterize the prior information of the images. The two submodules were combined to form the SLMD-Net method, and pre-clinical simulation data were used to validate the feasibility and effectiveness of the algorithm. RESULTS: Compared with the traditional model-driven quantitative imaging methods (FBP-DI, PWLS-PCG, and E3DTV), data-driven supervised-learning-based quantitative imaging methods (SUMD-Net and BFCNN), a material quantitative imaging method based on unsupervised learning (UNTV-Net) and semi-supervised learning-based cycle consistent generative adversarial network (Semi-CycleGAN), the proposed SLMD-Net method had better performance in both visual and quantitative assessments. For quantitative imaging of water and bone materials, the SLMD-Net method had the highest PSNR index (31.82 and 29.06), the highest FSIM index (0.95 and 0.90), and the lowest RMSE index (0.03 and 0.02), respectively) and achieved significantly higher image quality scores than the other 7 material decomposition methods (P < 0.05). The material quantitative imaging performance of SLMD-Net was close to that of the supervised network SUMD-Net trained with labeled data with a doubled size. CONCLUSIONS A small labeled dataset and a large unlabeled low SNR material image dataset can be fully used to suppress noise amplification and artifacts in basic material decomposition in spectral CT and reduce the dependence on labeled data-driven network, which considers more realistic scenario in clinics.
Tomography, X-Ray Computed/methods* ; Image Processing, Computer-Assisted/methods* ; Algorithms ; Signal-To-Noise Ratio ; Perception

OBJECTIVE: To propose a non-contrast CT-based algorithm for automated and accurate detection of pancreatic lesions at a low cost. METHODS: With Faster RCNN as the benchmark model, an advanced Faster RCNN (aFaster RCNN) model for pancreatic lesions detection based on plain CT was constructed. The model uses the residual connection network Resnet50 as the feature extraction module to extract the deep image features of pancreatic lesions. According to the morphology of pancreatic lesions, 9 anchor frame sizes were redesigned to construct the RPN module. A new Bounding Box regression loss function was proposed to constrain the training process of RPN module regression subnetwork by comprehensively considering the constraints of the lesion shape and anatomical structure. Finally, a detection frame was generated using the detector in the second stage. The data from a total of 728 cases of pancreatic diseases from 4 clinical centers in China were used for training (518 cases, 71.15%) and testing (210 cases, 28.85%) of the model. The performance of aFaster RCNN was verified through ablation experiments and comparison experiments with 3 classical target detection models SSD, YOLO and CenterNet. RESULTS: The aFaster RCNN model for pancreatic lesion detection achieved recall rates of 73.64% at the image level and 92.38% at the patient level, with an average precision of 45.29% and 53.80% at the image and patient levels, respectively, which were higher than those of the 3 models for comparison. CONCLUSION The proposed method can effectively extract the imaging features of pancreatic lesions from non-contrast CT images to detect the pancreatic lesions.
Humans ; Pancreas/diagnostic imaging* ; Algorithms ; China ; Pancreatic Neoplasms/diagnostic imaging* ; Tomography, X-Ray Computed

OBJECTIVE: To investigate the consistency and diagnostic performance of magnetic resonance imaging (MRI) for detecting microvascular invasion (MVI) of hepatocellular carcinoma (HCC) and the validity of deep learning attention mechanisms and clinical features for MVI grade prediction. METHODS: This retrospective study was conducted among 158 patients with HCC treated in Shunde Hospital Affiliated to Southern Medical University between January, 2017 and February, 2020. The imaging data and clinical data of the patients were collected to establish single sequence deep learning models and fusion models based on the EfficientNetB0 and attention modules. The imaging data included conventional MRI sequences (T₁WI, T₂WI, and DWI), enhanced MRI sequences (AP, PP, EP, and HBP) and synthesized MRI sequences (T1mapping-pre and T1mapping-20 min), and the high-risk areas of MVI were visualized using deep learning visualization techniques. RESULTS: The fusion model based on T1mapping-20min sequence and clinical features outperformed other fusion models with an accuracy of 0.8376, a sensitivity of 0.8378, a specificity of 0.8702, and an AUC of 0.8501 for detecting MVI. The deep fusion models were also capable of displaying the high-risk areas of MVI. CONCLUSION The fusion models based on multiple MRI sequences can effectively detect MVI in patients with HCC, demonstrating the validity of deep learning algorithm that combines attention mechanism and clinical features for MVI grade prediction.
Humans ; Carcinoma, Hepatocellular ; Retrospective Studies ; Liver Neoplasms ; Magnetic Resonance Imaging ; Algorithms

OBJECTIVE: To propose an deep learning-based algorithm for automatic prediction of dose distribution in radiotherapy planning for head and neck cancer. METHODS: We propose a novel beam dose decomposition learning (BDDL) method designed on a cascade network. The delivery matter of beam through the planning target volume (PTV) was fitted with the pre-defined beam angles, which served as an input to the convolution neural network (CNN). The output of the network was decomposed into multiple sub-fractions of dose distribution along the beam directions to carry out a complex task by performing multiple simpler sub-tasks, thus allowing the model more focused on extracting the local features. The subfractions of dose distribution map were merged into a distribution map using the proposed multi-voting mechanism. We also introduced dose distribution features of the regions-of-interest (ROIs) and boundary map as the loss function during the training phase to serve as constraining factors of the network when extracting features of the ROIs and areas of dose boundary. Public datasets of radiotherapy planning for head and neck cancer were used for obtaining the accuracy of dose distribution of the BDDL method and for implementing the ablation study of the proposed method. RESULTS: The BDDL method achieved a Dose score of 2.166 and a DVH score of 1.178 (P < 0.05), demonstrating its superior prediction accuracy to that of current state-ofthe-art (SOTA) methods. Compared with the C3D method, which was in the first place in OpenKBP-2020 Challenge, the BDDL method improved the Dose score and DVH score by 26.3% and 30%, respectively. The results of the ablation study also demonstrated the effectiveness of each key component of the BDDL method. CONCLUSION The BDDL method utilizes the prior knowledge of the delivery matter of beam and dose distribution in the ROIs to establish a dose prediction model. Compared with the existing methods, the proposed method is interpretable and reliable and can be potentially applied in clinical radiotherapy.
Humans ; Deep Learning ; Head and Neck Neoplasms/radiotherapy* ; Algorithms ; Neural Networks, Computer