Hip fractures pose a significant challenge to healthcare systems due to their high costs and associated mortality rates, with femoral neck fractures accounting for nearly half of all hip fractures. This study addresses the challenge of diagnosing nondisplaced femoral neck fractures, which are often difficult to detect with standard radiographs, especially in elderly patients. This research evaluates a deep learning model that employs a convolutional neural network (CNN) within a ResNet framework, designed to enhance diagnostic accuracy for nondisplaced femoral neck fractures. The model was trained and validated on a dataset of 2032 hip radiographs from two hospitals, with additional external validation performed on datasets from other institutions. The AI model achieved an accuracy of 94.8% and an Area Under Curve of 0.991 on anteroposterior pelvic/hip radiographs, outperforming emergency physicians and delivering results comparable to expert physicians. External validation confirmed the model’s robust accuracy and generalizability across diverse datasets. This study underscores the potential of deep learning models to act as a supplementary tool in clinical settings, potentially reducing diagnostic errors and improving patient outcomes by facilitating a quicker diagnosis and treatment.
Background: During transcatheter arterial chemoembolization (TACE), the delivery of a guidewire to the lesion site is a critical step, making the analysis and positioning of guidewire morphology crucial for both robotic systems and physicians in interventional surgeries. Current research on guidewires often faces challenges such as a low image signal-to-noise ratio and severe class imbalance. To overcome these issues and enhance the practical delivery of guidewires in clinical settings, this study introduces a comprehensive dataset for guidewire delivery during TACE and develops a specialized deep learning model for segmenting guidewire morphology in X-ray fluoroscopic images. Methods: We retrospectively collected 2,839 X-ray images acquired under real-time guidance from 38 subjects and manually annotated the guidewires. We proposed a deep learning-based guidewire segmentation method, which integrated two effective modules designed in this study: a bilateral feature fusion (BGA) module and a lightweight gated attention (SDA) module, achieving precise segmentation of guidewires in intraoperative images. Results: Quantitative and qualitative assessments were performed on 903 clinical images from 27 X-ray fluoroscopy videos. The segmentation network proposed in this paper demonstrated superior performance, achieving an area under the curve (AUC) of 91.64%, a Macro-F1 score of 85.63%, and a Dice coefficient of 71.29%. Conclusions: This study introduces a novel guidewire segmentation method specifically designed for clinical TACE. It not only assists physicians during interventional procedures but is also expected to be integrated into the intelligent systems of vascular interventional surgical robots, enabling robotic assistance in the future interventional surgeries.