Chi-Yuan Wang

Pulmonary embolism (PE) on CT pulmonary angiography (CTPA) demands rapid, accurate assessment, yet small, low-contrast clots in distal arteries remain challenging. We benchmarked ten fully convolutional network (FCN) backbones and introduced Consensus Intersection-Optimized Fusion (CIOF)—a K-of-M, pixel-wise mask fusion with the voting threshold K* selected on training patients to maximize IoU. Using the FUMPE cohort (35 patients; 12,034 slices) with patient-based random splits (18 train, 17 test), we trained five FCN architectures (each with Adam and SGDM) and evaluated segmentation with IoU, Dice, FNR/FPR, and latency. CIOF achieved the best overall performance (mean IoU 0.569; mean Dice 0.691; FNR 0.262), albeit with a higher runtime (~63.7 s per case) because all ten models are executed and fused; the strongest single backbone was Inception-ResNetV2 + SGDM (IoU 0.530; Dice 0.648). Stratified by embolization ratio, CIOF remained superior across 10⁻³ clot burdens, with mean IoU/Dice = 0.238/0.328, 0.566/0.698, and 0.739/0.846, respectively—demonstrating gains for tiny, subsegmental emboli. These results position CIOF as an accuracy-oriented, interpretable ensemble for offline or second-reader use, while faster single backbones remain candidates for time-critical triage.

Pulmonary embolism (PE) is a critical cardiovascular condition requiring rapid and accurate detection. This study presents an AI-enhanced biosensing framework utilizing computed tomography angiography (CTA) as an imaging-based sensor platform for automated PE diagnosis. Deep learning-based semantic segmentation models were employed to extract pathophysiological signals indicative of emboli. We utilized the publicly available FUMPE dataset, containing 2,304 annotated CTA slices from 35 patients. Ten fully convolutional networks (FCNs) were trained using five backbone architectures each optimized with both Adam and SGDM algorithms. To improve robustness and mitigate model-specific biases, we propose a novel Consensus Intersection-Optimized Fusion (CIOF) method, which adaptively aggregates multi-model predictions to maximize overlap with ground truth. CIOF achieved only 41 missed detections, outperforming all individual models (e.g., InceptionResNetV2+Adam: 411 misses). Performance was evaluated using IoU, Dice coefficient, boundary-based F1 score, and the number of segmented regions. The CIOF mechanism emulates intelligent signal fusion in biosensing applications and supports real-time integration with diagnostic imaging platforms. This work highlights CIOF as a promising AI-driven ensemble fusion strategy for medical biosensing, offering potential for integration into smart diagnostic systems, computer-aided tools, and future wearable biosensor-assisted triage technologies.

Background/Objectives: Breast cancer is a leading cause of mortality among women in Taiwan and globally. Non-invasive imaging methods, such as mammography and ultrasound, are critical for early detection, yet standalone modalities have limitations in regard to their diagnostic accuracy. This study aims to enhance breast cancer detection through a cross-modality fusion approach combining mammography and ultrasound imaging, using advanced convolutional neural network (CNN) architectures. Materials and Methods: Breast images were sourced from public datasets, including the RSNA, the PAS, and Kaggle, and categorized into malignant and benign groups. Data augmentation techniques were used to address imbalances in the ultrasound dataset. Three models were developed: (1) pre-trained CNNs integrated with machine learning classifiers, (2) transfer learning-based CNNs, and (3) a custom-designed 17-layer CNN for direct classification. The performance of the models was evaluated using metrics such as accuracy and the Kappa score. Results: The custom 17-layer CNN outperformed the other models, achieving an accuracy of 0.964 and a Kappa score of 0.927. The transfer learning model achieved moderate performance (accuracy 0.846, Kappa 0.694), while the pre-trained CNNs with machine learning classifiers yielded the lowest results (accuracy 0.780, Kappa 0.559). Cross-modality fusion proved effective in leveraging the complementary strengths of mammography and ultrasound imaging. Conclusions: This study demonstrates the potential of cross-modality imaging and tailored CNN architectures to significantly improve diagnostic accuracy and reliability in breast cancer detection. The custom-designed model offers a practical solution for early detection, potentially reducing false positives and false negatives, and improving patient outcomes through timely and accurate diagnosis.

This study focuses on overcoming challenges in classifying eye diseases using color fundus photographs by leveraging deep learning techniques, aiming to enhance early detection and diagnosis accuracy. We utilized a dataset of 6392 color fundus photographs across eight disease categories, which was later augmented to 17,766 images. Five well-known convolutional neural networks (CNNs)—efficientnetb0, mobilenetv2, shufflenet, resnet50, and resnet101—and a custom-built CNN were integrated and trained on this dataset. Image sizes were standardized, and model performance was evaluated via accuracy, Kappa coefficient, and precision metrics. Shufflenet and efficientnetb0demonstrated strong performances, while our custom 17-layer CNN outperformed all with an accuracy of 0.930 and a Kappa coefficient of 0.920. Furthermore, we found that the fusion of image features with classical machine learning classifiers increased the performance, with Logistic Regression showcasing the best results. Our study highlights the potential of AI and deep learning models in accurately classifying eye diseases and demonstrates the efficacy of custom-built models and the fusion of deep learning and classical methods. Future work should focus on validating these methods across larger datasets and assessing their real-world applicability.

According to the Health Promotion Administration in the Ministry of Health and Welfare statistics in Taiwan, over ten thousand women have breast cancer every year. Mammography is widely used to detect breast cancer. However, it is limited by the operator’s technique, the cooperation of the subjects, and the subjective interpretation by the physician. It results in inconsistent identification. Therefore, this study explores the use of a deep neural network algorithm for the classification of mammography images. In the experimental design, a retrospective study was used to collect imaging data from actual clinical cases. The mammography images were collected and classified according to the breast image reporting and data-analyzing system (BI-RADS). In terms of model building, a fully convolutional dense connection network (FC-DCN) is used for the network backbone. All the images were obtained through image preprocessing, a data augmentation method, and transfer learning technology to build a mammography image classification model. The research results show the model’s accuracy, sensitivity, and specificity were 86.37%, 100%, and 72.73%, respectively. Based on the FC-DCN model framework, it can effectively reduce the number of training parameters and successfully obtain a reasonable image classification model for mammography.

Many neurological and musculoskeletal disorders are associated with problems related to postural movement. Noninvasive tracking devices are used to record, analyze, measure, and detect the postural control of the body, which may indicate health problems in real time. A total of 35 young adults without any health problems were recruited for this study to participate in a walking experiment. An iso-block postural identity method was used to quantitatively analyze posture control and walking behavior. The participants who exhibited straightforward walking and skewed walking were defined as the control and experimental groups, respectively. Fusion deep learning was applied to generate dynamic joint node plots by using OpenPose-based methods, and skewness was qualitatively analyzed using convolutional neural networks. The maximum specificity and sensitivity achieved using a combination of ResNet101 and the naïve Bayes classifier were 0.84 and 0.87, respectively. The proposed approach successfully combines cell phone camera recordings, cloud storage, and fusion deep learning for posture estimation and classification.

Positron emission tomography (PET) can provide functional images and identify abnormal metabolic regions of the whole-body to effectively detect tumor presence and distribution. The filtered back-projection (FBP) algorithm is one of the most common images reconstruction methods. However, it will generate strike artifacts on the reconstructed image and affect the clinical diagnosis of lesions. Past studies have shown reduction in strike artifacts and improvement in quality of images by two-dimensional morphological structure operators (2D-MSO). The morphological structure method merely processes the noise distribution of 2D space and never considers the noise distribution of 3D space. This study was designed to develop three-dimensional-morphological structure operators (3D MSO) for nuclear medicine imaging and effectively eliminating strike artifacts without reducing image quality. A parallel operation was also used to calculate the minimum background standard deviation of the images for three-dimensional morphological structure operators with the optimal response curve (3D-MSO/ORC). As a result of Jaszczak phantom and rat verification, 3D-MSO/ORC showed better denoising performance and image quality than the 2D-MSO method. Thus, 3D MSO/ORC with a 3 × 3 × 3 mask can reduce noise efficiently and provide stability in FBP images.

Anesthesia assessment is most important during surgery. Anesthesiologists use electrocardiogram (ECG) signals to assess the patient’s condition and give appropriate medications. However, it is not easy to interpret the ECG signals. Even physicians with more than 10 years of clinical experience may still misjudge. Therefore, this study uses convolutional neural networks to classify ECG image types to assist in anesthesia assessment. The research uses Internet of Things (IoT) technology to develop ECG signal measurement prototypes. At the same time, it classifies signal types through deep neural networks, divided into QRS widening, sinus rhythm, ST depression, and ST elevation. Three models, ResNet, AlexNet, and SqueezeNet, are developed with 50% of the training set and test set. Finally, the accuracy and kappa statistics of ResNet, AlexNet, and SqueezeNet in ECG waveform classification were (0.97, 0.96), (0.96, 0.95), and (0.75, 0.67), respectively. This research shows that it is feasible to measure ECG in real time through IoT and then distinguish four types through deep neural network models. In the future, more types of ECG images will be added, which can improve the real-time classification practicality of the deep model.

Postural control decreases with aging. Thus, an efficient and accurate method of detecting postural control is needed. We enrolled 35 elderly adults (aged 82.06 ± 8.74 years) and 20 healthy young adults (aged 21.60 ± 0.60 years) who performed standing tasks for 40 s, performed six times. The coordinates of 15 joint nodes were captured using a Kinect device (30 Hz). We plotted joint positions into a single 2D figure (named a joint–node plot, JNP) once per second for up to 40 s. A total of 15 methods combining deep and machine learning for postural control classification were investigated. The accuracy, sensitivity, specificity, positive predicted value (PPV), negative predicted value (NPV), and kappa values of the selected methods were assessed. The highest PPV, NPV, accuracy, sensitivity, specificity, and kappa values were higher than 0.9 in validation testing. The presented method using JNPs demonstrated strong performance in detecting the postural control ability of young and elderly adults.