Implementation of an Attention Mechanism Model for Facial Beauty Assessment Using Transfer Learning.

An important consideration in medical plastic surgery is the evaluation of the patient's facial symmetry. However, because facial attractiveness is a slightly individualized cognitive experience, it is difficult to determine face attractiveness manually. This study aimed to train a model for assessing facial attractiveness using transfer learning while also using the fine-grained image model to separate similar images by first learning features. In this case, the system can make assessments based on the input of facial photos. Thus, doctors can quickly and objectively treat patients' scoring and save time for scoring. The transfer learning was combined with CNN, Xception, and attention mechanism models for training, using the SCUT-FBP5500 dataset for pre-training and freezing the weights as the transfer learning model. Then, we trained the Chang Gung Memorial Hospital Taiwan dataset to train the model based on transfer learning. The evaluation uses the mean absolute error percentage (MAPE) value. The root mean square error (RMSE) value is used as the basis for experimental adjustment and the quantitative standard for the model's predictive. The best model can obtain 0.50 in RMSE and 18.5% average error in MAPE. A web page was developed to infer the deep learning model to visualize the predictive model.

Electrically programmable magnetoresistance in [Formula: see text]-based magnetic tunnel junctions.

We report spin-dependent transport properties and I-V hysteresis characteristics in an [Formula: see text]-based magnetic tunnel junction (MTJ). The bipolar resistive switching and the magnetoresistances measured at high resistance state (HRS) and low resistance state (LRS) yield four distinctive resistive states in a single device. The temperature dependence of resistance at LRS suggests that the resistive switching is not triggered by the metal filaments within the [Formula: see text] layer. The role played by oxygen vacancies in [Formula: see text] is the key to determine the resistive state. Our study reveals the possibility of controlling the multiple resistive states in a single [Formula: see text]-based MTJ by the interplay of both electric and magnetic fields, thus providing potential applications for future multi-bit memory devices.

On construction of transfer learning for facial symmetry assessment before and after orthognathic surgery.

Orthognathic surgery (OGS) is frequently used to correct facial deformities associated with skeletal malocclusion and facial asymmetry. An accurate evaluation of facial symmetry is a critical for precise surgical planning and the execution of OGS. However, no facial symmetry scoring standard is available. Typically, orthodontists or physicians simply judge facial symmetry. Therefore, maintaining accuracy is difficult. We propose a convolutional neural network with a transfer learning approach for facial symmetry assessment based on 3-dimensional (3D) features to assist physicians in enhancing medical treatments. We trained a new model to score facial symmetry using transfer learning. Cone-beam computed tomography scans in 3D were transformed into contour maps that preserved 3D characteristics. We used various data preprocessing and amplification methods to determine the optimal results. The original data were enlarged by 100 times. We compared the quality of the four models in our experiment, and the neural network architecture was used in the analysis to import the pretraining model. We also increased the number of layers, and the classification layer was fully connected. We input random deformation data during training and dropout to prevent the model from overfitting. In our experimental results, the Xception model and the constant data amplification approach achieved an accuracy rate of 90%.

Spin pump and probe in lanthanum strontium manganite/platinum bilayers.

Ferromagnetic resonance driven spin pumping (FMR-SP) is a novel method to transfer spin current from the ferromagnetic (FM) layer into the adjacent normal metal (NM) layer in an FM/NM bilayer system. Consequently, the spin current could be probed in NM layer via inverse spin Hall effect (ISHE). In spite of numerous ISHE studies on FM/Pt bilayers, La0.7Sr0.3MnO3(LSMO)/Pt system has been less explored and its relevant information about interface property (characterized by spin mixing conductance) and spin-charge conversion efficiency (characterized by spin Hall angle) is a matter of importance for the possible applications of spintronic devices. In this work, the technique of FMR-SP has been applied on two series of LSMO/Pt bilayers with the thickness of each layer being varied. The thickness dependences of ISHE voltage allow to extract the values of spin mixing conductance and spin Hall angle of LSMO/Pt bilayers, which are (1.8 ± 0.4) × 1019 m-2 and (1.2 ± 0.1) % respectively. In comparison with other FM/Pt systems, LSMO/Pt has comparable spin current density and spin mixing conductance, regardless its distinct electronic structure from other ferromagnetic metals.

A new method for the integration of digital dental models and cone-beam computed tomography images.

This study introduces a regional-surface-based registration without markers for integration of laser-scanned dental images into maxillofacial cone-beam computed tomography (CBCT) images. The method just needs to manually select three similar areas without artifact on the digital dental image and CBCT image, and then the process is automatically complete the fusion (superimposition) of maxillofacial model and the digital dental model. Then the differences such as mean error and root-mean-square (RMS) error are automatically computed between the 2 images according to the selected surfaces and expressed in a color scale. Experimental results show that the mean errors between the 2 models at the integrated model range from 0.15 mm to 0.45 mm and the RMS errors range 0.18 mm to 0.49 mm. The numbers are similar to the results of previous methods and reach a desirable error. Moreover, it is robust feasibility for especially serious artifacts CBT images. It is worth mentioning that all measurements of intra-operator reproducibility and inter-operator reliability are excellent.

Artifact-resistant superimposition of digital dental models and cone-beam computed tomography images.

PURPOSE: Combining the maxillofacial cone-beam computed tomography (CBCT) model with its corresponding digital dental model enables an integrated 3-dimensional (3D) representation of skeletal structures, teeth, and occlusions. Undesired artifacts, however, introduce difficulties in the superimposition of both models. We have proposed an artifact-resistant surface-based registration method that is robust and clinically applicable and that does not require markers. MATERIALS AND METHODS: A CBCT bone model and a laser-scanned dental model obtained from the same patient were used in developing the method and examining the accuracy of the superimposition. Our method included 4 phases. The first phase was to segment the maxilla from the mandible in the CBCT model. The second phase was to conduct an initial registration to bring the digital dental model and the maxilla and mandible sufficiently close to each other. Third, we manually selected at least 3 corresponding regions on both models by smearing patches on the 3D surfaces. The last phase was to superimpose the digital dental model into the maxillofacial model. Each superimposition process was performed twice by 2 operators with the same object to investigate the intra- and interoperator differences. All collected objects were divided into 3 groups with various degrees of artifacts: artifact-free, critical artifacts, and severe artifacts. The mean errors and root-mean-square (RMS) errors were used to evaluate the accuracy of the superimposition results. Repeated measures analysis of variance and the Wilcoxon rank sum test were used to calculate the intraoperator reproducibility and interoperator reliability. RESULTS: Twenty-four maxilla and mandible objects for evaluation were obtained from 14 patients. The experimental results showed that the mean errors between the 2 original models in the residing fused model ranged from 0.10 to 0.43 mm and that the RMS errors ranged from 0.13 to 0.53 mm. These data were consistent with previously used methods and were clinically acceptable. All measurements of the proposed study exhibited desirable intraoperator reproducibility and interoperator reliability. Regarding the intra- and interoperator mean errors and RMS errors in the nonartifact or critical artifact group, no significant difference between the repeated trials or between operators (P < .05) was observed. CONCLUSIONS: The results of the present study have shown that the proposed regional surface-based registration can robustly and accurately superimpose a digital dental model into its corresponding CBCT model.