Abstract: Abstract: To address the issue of insufficient learning of features between different organs in the segmentation of pheochromocytoma images from partially annotated abdominal datasets, which leads to difficulties in accurately distinguishing tumor and surrounding organ boundaries, a Multi-scale and Spatial Frequency feature-based image segmentation Network for pheochromocytoma (MFC-Net) was proposed. Firstly, a Multi-scale Spatial Frequency Channel Attention module (MSFCA) was constructed, which enhanced the capture of inter-organ texture and boundary features by weighting and fusing the frequency domain information of the image and multi-scale feature maps from adjacent encoders, thereby highlighting the feature representation of tumor regions. Then, an Up-sampling Multi-Scale Feature Fusion module (UMFF) was introduced, which combined up-sampled feature maps of different scales to enhance the model's ability to recognize objects of varying sizes in the image. Finally, an Adaptive Objective Loss Function (AOb) was utilized, which calculated the loss based on annotated abdominal organ labels and adjusted the loss weights according to the annotated organ categories, optimizing the learning process of the segmentation network. Experimental results showed that on the abdominal organ and pheochromocytoma datasets, the segmentation accuracy of MFC-Net improved by 3.33 and 3.18 percentage points compared to separately trained nnU-Net, with Dice and NSD scores of 89.07% and 92.85%, respectively. When tested on out-of-domain datasets, the scores were 84.66% and 90.55%. Visualization results indicated that MFC-Net could better handle complex backgrounds and blurred boundaries in pheochromocytoma images, providing better technical support for the accurate diagnosis and treatment of pheochromocytoma.

Key words: Keywords: abdominal organ segmentation, pheochromocytoma medical image, Fourier transform, frequency domain processing, partially labeled dataset

摘要： 摘 要: 针对腹部部分注释数据集嗜铬细胞瘤图像分割缺乏不同器官间的特征的学习，导致分割中难以准确区分肿瘤及周边器官边缘的问题，提出了一种基于多尺度与空间频率特征的嗜铬细胞瘤图像分割方法（MFC-Net）。首先，构建多尺度空间频率通道注意力模块（MSFCA），通过对图像频域信息和相邻编码器多尺度特征图进行加权融合，强化器官间纹理和边界特征的捕捉，突出肿瘤区域的特征表示能力；其次，多尺度特征融合模块（UMFF），结合上采样得到的不同尺度特征图，增强模型对图像中不同大小对象的识别能力；最后，利用自适应目标损失函数（AOb），针对有注释腹部器官标签进行损失计算，根据注释器官类别调整损失权重大小，优化分割网络的学习过程。实验结果表明，在腹部器官和嗜铬细胞瘤数据集上，MFC-Net分割准确率相较于单独训练nnU-Net分别提升了3.33个百分点和3.18个百分点，Dice和NSD分别为89.07%和92.85%，在域外数据集上测试为84.66%和90.55%，且可视化结果表明能够更好的处理嗜铬细胞瘤图像中的复杂背景和模糊边界，为嗜铬细胞瘤的精确诊断和治疗提供更好的技术支持。

关键词: 关键词: 腹部器官分割, 嗜铬细胞瘤医学图像, 傅里叶变换, 频域处理, 部分标签数据集