Research on Obstructed Object Detection Algorithms in Medical X-ray images Based on Deep Learning

In this paper explores the optimization of object detection algorithms for occlusions in X-ray imaging. Addressing the absence of case segmentation datasets for nursing scenarios, we compiled a substantial collection of clear and accurately labeled image datasets. After preprocessing the data, which included image scaling, linear transformation, and contrast enhancement, we conducted target detection experiments using SSD, Faster R-CNN, and optimized Faster R-CNN models. The evaluation of the models was based on the regression accuracy calculated by comparing the predicted and actual marked positions. The results indicate that while the SSD model is capable of performing occlusion area detection tasks, it is not ideal, with 11 instances of false detection and significant deviation. The traditional Faster R-CNN model showed improvement across various metrics, reducing the number of false detections to 3, yet the accuracy remained suboptimal. In contrast, the optimized Faster R-CNN model, which integrates a feature pyramid network, utilizes ResNet for feature extraction, and transforms the size, shape, position, and boundary polarity of the input data’s occlusion area, achieved stable loss values more rapidly. It demonstrated high detection accuracy with intersection rates of 74.8% and 79.0% for single-location and dual-location occlusion detection, respectively. These findings underscore the method’s feasibility and potential for practical application.