This study aimed to evaluate whether the hybrid and biomechanically deformable image registration (DIR) algorithm of the RayStation treatment planning system would produce contour propagation and dose deformation errors in the head and neck due to the inclusion of adjuvant therapeutic fixtures. We analyzed the treatment plans of two groups of patients with head and neck cancer (Planx and Planp). Planx included photon beam therapy (5250cGy/25 sessions) and Planp involved proton beam therapy (1680cGy/8 sessions). We used two adjuvant treatment immobilization devices (immobilization) to scan computed tomography (CT) images: Planx included CTx and immobilizationx, and Planp included CTp and immobilizationp. Using the hybrid (Hy) and biomechanical (Bio) algorithms of the RayStation treatment planning system, we controlled the registration to analyze the contour propagation and dose deformation. The range of immobilization including the body contour is defined as Rim+b, and the range of only the body contour is defined as Rb. We generated four settings as follows: Hy_Rim+b, Bio_Rim+b, Hy_Rb, and Bio_Rb. We mapped organs at risk (OARs) to Planp by using contour propagation through the aforementioned four settings. Contour propagation uses the results of overlapping image display, the Dice similarity coefficient (DSC), and the contour drawn by the physician on Planp. We used the results shown in the overlapping images in the contour propagation and evaluated them with the DSC and the contour drawn by the physician in Planp. We mapped the received dose of OARs in Planx to Planp with dose deformation, and evaluated the percent dose difference [dose diff.(%)] between the four settings and Planx. In terms of contour propagation, the overlapping image of the horizontal section (transversal) showed that because the range set by Hy_Rim+b and Bio_Rim+b includes immobilization, Hy_Rim+b deforms in the oral cavity and esophagus area, and for Bio_Rim+b significant deformations around the body contour lead to misregistration. The Hy_Rb and Bio_Rb settings are not obviously deformed in the overlapping images. We assessed the consistency of dissemination of OARs contours by using the DSC. The average DSC of Hy_Rim+b and Bio_Rim+b is 0.63 and 0.32, respectively; the average DSC of Hy_Rb and Bio_Rb is 0.94 and 0.83, respectively. The results of the overlapping image and DSC evaluation showed that the two algorithms can reduce the error by excluding immobilization in the registration range of contour propagation. We found that the hybrid algorithm is superior to the biomechanical algorithm. In terms of dose deformation, the average dose differences of Hy_Rim+b and Bio_Rim+b in Planx are 13.15% and 17.82%, respectively, while those of Hy_Rb and Bio_Rb are 3.32% and 5.13%, respectively. We found that the average dose error of the hybrid algorithm is smaller than that of the biomechanical algorithm. Considering the setting where the registration range does or does not include immobilization, the average dose of OARs differs by 9.83% for the hybrid algorithm and 12.69% for the biomechanical algorithm. In conclusion, we found that the hybrid and biomechanical algorithms of the RayStation treatment planning system increase the error of contour propagation and dose deformation because the registration range includes head and neck immobilization. The results show that the hybrid algorithm is more suitable for the head and neck than the biomechanical algorithm. Therefore, we suggest using the hybrid algorithm for clinical planning of DIR, and excluding immobilization from taking the patient's body contour as the registration range.

Using machine learning algorithms to analyze the odds and predictors of complications of thyroid damage after radiation therapy in patients with head and neck cancer. This study used decision tree (DT), random forest (RF), and support vector machine (SVM) algorithms to evaluate predictors for the data of 137 head and neck cancer patients. Candidate factors included gender, age, thyroid volume, minimum dose, average dose, maximum dose, number of treatments, and relative volume of the organ receiving X dose (X: 10, 20, 30, 40, 50, 60 Gy). The algorithm was optimized according to these factors and cross-validation was performed 10 times to analyze the state of thyroid damage and select the predictors of thyroid dysfunction. The importance of the predictors identified by the three machine learning algorithms was ranked: the top five predictors were age, thyroid volume, average dose, V50 and V60. Of these, age and volume were negatively correlated with thyroid damage, indicating that the greater the age and thyroid volume, the lower the risk of thyroid damage; the average dose, V50 and V60 were positively correlated with thyroid damage, indicating that the larger the average dose, V50 and V60, the higher the risk of thyroid damage. The RF algorithm was most accurate in predicting the probability of thyroid damage among the three algorithms optimized using the above factors. The AUC was 0.827 and the ACC was 0.824. This study found that five predictors (age, thyroid volume, mean dose, V50 and V60) are important factors affecting the chance that patients with head and neck cancer who received radiation therapy will develop hypothyroidism. Using these factors as the prediction basis of the algorithm and using RF to predict the occurrence of hypothyroidism had the highest accuracy, which was 82.4%. This algorithm can be used as a reference for predicting the probability of radiation therapy complications and assisting medical decision-making in the future.

To establish a suitable deep learning model in order to predict the risk of diabetic Peripheral Neuropathy (DPN) by using fundus photography images of type II diabetic patients with artificial intelligence (AI) approach. From the year 2013 to 2017, by using a diabetes care database established by the Department of Endocrinology and Metabolism of Kaohsiung Datong Hospital (KMTTH) and the Affiliated Hospital of Kaohsiung Medical University (KMUH), a corresponding retrospective study of fundus photographic images of patients with type II diabetes mellitus was performed. Patients with Type II diabetes who have undergone clinical routine treatment and have routine ophthalmoscopy were analyzed and classified according to the results of the Nerve Conduction Velocity (NCV) method. Soon after the patient's personal information was removed, the image is preprocessed by adaptive histogram equalization (CLAHE) to limit the contrast variation. These preprocessed images were then divided into training, validation, and test sets. Another two sets of Rotated image data were also incorporated for enhancements to build prediction models through four deep learning architectures: InceptionNet, VGGNet, ResNet, and ConvMixer DPN model, respectively. In this study, a classification model for predicting the severity of DPN was successfully established through four deep learning architectures. The accuracies of the four DPN prediction models established in this study were 0.94, 0.90, 0.97, and 0.96; AUC values have achieved at 0.92, 0.93, 0.95, and 0.96; specificity analyses were 1.00, 0.92, 1.00, and 0.98; The combined sensitivity values of mild and moderate to severe DPN reached 0.84, 0.90, 0.90, and 0.92, respectively. An AI-assisted diagnostic model was successfully established to predict the severity of diabetic peripheral neuropathy (DPN), which could determine whether the patient has DPN from the retinal fundus images obtained after ophthalmoscopy and with its associated severity; therefore, it is an efficient, non-invasive method of DPN detection.

Helicobacter pylori infection is the principal cause of chronic gastritis, gastric ulcers, duodenal ulcers, and gastric cancer. In clinical practice, diagnosis of H. pylori infection by a gastroenterologists’ impression of endoscopic images is inaccurate and cannot be used for the management of gastrointestinal diseases. The aim of this study was to develop an artificial intelligence classification system for the diagnosis of H. pylori infection by pre-processing endoscopic images and machine learning methods. Endoscopic images of the gastric body and antrum from 302 patients receiving endoscopy with confirmation of H. pylori status by a rapid urease test at An Nan Hospital were obtained for the derivation of an artificial intelligence classification system. The H. pylori status was interpreted as positive or negative by CNN and Concurrent Spatial and Channel Squeeze and Excitation (scSE), combined with different classification models for deep learning of gastric images. The comprehensive assessment for H. pylori status by scSE-CatBoost classification models for both body and antrum images achieved an accuracy of 0.90, sensitivity of 1.00, specificity of 0.81, positive predictive value of 0.82, negative predicted value of 1.00, and area under the curve of 0.88. The data suggest that an artificial intelligence classification model using scSE’s deep learning for gastric endoscopic images can distinguish H. pylori status with good performance and is useful for the survey or diagnosis of H. pylori infection in clinical practice.