Weak redundancy U-shaped network and heatmap-based object prompt method for real-time medical image processing

In recent years, artificial intelligence (AI) technology has experienced significant growth, leading to the development of advanced tools that assist radiologists in image interpretation and diagnostic decision-making. In the field of medical image processing, object detection and segmentation are crucial research areas. Achieving automatic and rapid segmentation of organs and lesions can significantly enhance physicians' work efficiency. However, most existing networks feature complex architectures and entail high computational complexity. Due to the limited processing power of the devices, achieving real-time segmentation often proves challenging. Therefore, the aim of this study was to design a lightweight segmentation algorithm that reduces resource consumption and enhances real-time performance.

Firstly, we propose a compact U-shaped network called the weak redundancy U-Net (WRU-Net) specifically designed for real-time segmentation tasks in medical imaging. By reducing the number of channels, feature redundancy across all scales is reduced, compelling the network to utilize resources efficiently at every level. Furthermore, we propose "auxiliary information flows" to facilitate the propagation of phased results, thereby enhancing the decoder. Secondly, this paper introduces a novel visual prompt mode that differs from the standard object detection mode. We refer to this as "object prompt", which means visualizing the position of the target object in an image to guide the viewer. Unlike standard object detection tasks that provide bounding boxes, this paper achieves the aforementioned effect in the form of heatmaps. Correspondingly, we propose a network for heatmap prediction, which further simplifies task complexity and achieves semantic detection of the target object.

We conducted experiments using the datasets of the thyroid nodule, chest, placental vessel, brain tumors, heart, liver, and spleen, which include medical images of X-ray, computed tomography (CT), and ultrasound. Furthermore, we conducted comparative experiments using several state-of-the-art networks. The segmentation performance of the networks was evaluated using metrics such as the Dice similarity coefficient (DSC), intersection over union (IoU), precision, and recall. The average DSC, IoU, precision, and recall of our model across each dataset were 88.53%, 82.58%, 86.34%, and 84.85%, respectively. Regarding efficiency, our WRU-Net achieved 130.67 KB model size and 488 frames per second (FPS), outperforming larger models. For heatmap prediction, our network exhibited a similarly efficient profile with a parameter size of only 24.75 KB and a speed of 9,862.13 FPS on graphics processing unit (GPU) and 107.93 FPS on central processing unit (CPU).

In scenarios where the calculation ability of the equipment is constrained or real-time performance requirements are stringent, the design of lightweight networks serves as a fundamental approach to achieving high-speed data stream processing functionality. For instance, this applies to embedded devices and real-time ultrasonic scanning applications. The method proposed in this paper significantly reduces the number of parameters, thereby decreasing computational resource consumption and enhancing speed. This improvement holds substantial significance for optimizing computational efficiency in practical applications.