Artificial Intelligence Analysis for Small Object Detection in Urine Sediment Images

Over recent years, artificial intelligence (AI) methods have reformed the zone of medical imaging. Many AI-based models have been created and improved to related medical image analysis and interpretation. Object detection in medical images is a fundamental task in computer vision. With the continuous development of deep learning technology in recent years, we have been able to achieve good detection performance for conventional objects, but it has still been challenging for small objects due to their small size. In this paper, we propose an artificial intelligence analysis method for object detection in urine sediment images for auxiliary diagnosis. Firstly, we introduce a New Path Aggregation Network (NPANet), which improves the multi-scale fusion section. It can be more beneficial to help detect relatively small cellular objects, thus reducing the rate of missed detection. Secondly, we design an Adaptive-IoU ( $AIoU$ ) loss that can be more easily adopted to achieve better detection accuracy for the model. Finally, the experiments are conducted on our Dataset (named UriSed2K) and the public Urine Microscopic Image Dataset, and achieve the state-of-the-art results. Especially, our detector reaches 61.8% on $AP_{small}$ on the UriSed2K and our work also contributes to the development of medical image processing.