基于U-Net#网络的前列腺肿瘤病理图像分割模型研究
Research on Prostate Tumor Pathology Image Segmentation Model Based on U-Net# Network

目的：本研究旨在分析基于U-Net#网络的分割模型对前列腺肿瘤病理图像分割的价值。方法：在经典U-Net架构的基础上，结合密集跳跃连接和全尺度跳跃连接，整合不同尺度的特征图，更好地捕捉前列腺肿瘤病理图像的语义特征。通过对上海交通大学医学院附属仁济医院提供的数据集进行扩增和预处理，构建了包含多种病理变化的前列腺肿瘤图像数据集。在训练过程中，采用Dice相关系数(dice similarity coefficient, DSC)和IoU值(Intersection over Union)作为评价指标，评估模型的分割性能。结果：通过对不同类UNet网络进行实验对比，U-Net#模型在前列腺肿瘤病理图像分割任务中实现了最高的DSC (85.43%)和IoU值(74.27%)。结论：实验结果表明，U-Net#中重新设计的跳跃连接成功地融合了多尺度上下文语义信息。与大多数最先进的医学图像分割模型相比，本文提出的方法更准确地定位器官和病变并精确分割边界。
Objective: This study aims to analyze the value of a segmentation model based on U-Net# network in the segmentation of pathological images of prostate tumors. Method: Based on the classic U-Net architecture, combining dense skip connections and full-scale skip connections, integrating feature maps of different scales to better capture the semantic features of prostate tumor pathological images. A prostate tumor image dataset containing multiple pathological changes was constructed by amplifying and preprocessing the dataset provided by Renji Hospital affiliated with Shanghai Jiao Tong University School of Medicine. During the training process, Dice similarity coefficient (DSC) and Intersection over Union (IoU) are used as evaluation metrics to assess the segmentation performance of the model. Result: Through experimental comparison of different types of UNet networks, the U-Net# model achieved the highest DSC (85.43%) and IoU values (74.27%) in the task of prostate tumor pathological image segmentation. Conclusion: The experimental results indicate that the redesigned skip connections in UNet# successfully integrate multi-scale contextual semantic information. Compared with most advanced medical image segmentation models, the method proposed in this paper can more accurately locate organs and lesions and accurately segment boundaries.