Our research team addresses the challenges in knee cartilage segmentation from magnetic resonance images, where deep convolutional neural networks (CNNs) often struggle with small samples of irregular shapes and tend to omit important features such as cartilage, despite their promising accuracy improvements, by proposing a novel network that incorporates a Patch Attention (PA) block and a Feature Aggregation block. The PA block consists of a Patch-based Channel-wise Attention block and a Patch-based Patch-wise Attention block, which capture intra-channel and intra-patch relationships, respectively, to enhance the network's ability to detect small objects. The Feature Aggregation block fuses features from the same level of the encoder and the previous layer of the decoder to enrich feature representation. We evaluate our method on two publicly available datasets, the 2010 Grand Challenge Knee Image Segmentation (SKI-10) dataset and the Osteoarthritis Initiative (OAI) dataset, and the results demonstrate that our approach achieves impressive performance in knee cartilage segmentation.

    研究团队针对膝关节磁共振图像软骨分割中，深度卷积神经网络尽管在精度提升上展现出潜力，但面对不规则形状的小样本时常常力不从心，容易遗漏软骨等重要特征的问题，提出了一种新型分割网络。该网络引入了Patch注意力（PA，Patch Attention）模块和特征聚合模块：PA模块包含基于补丁的通道注意力子块和基于补丁的块内注意力子块，分别捕获通道内和补丁内的关系，以增强网络对小目标的检测能力；特征聚合模块则融合编码器同层特征和解码器前一层的特征，以丰富特征表达。团队在2010年大挑战膝关节图像分割（SKI-10，the 2010 Grand Challenge Knee Image Segmentation）数据集和骨关节炎倡议（OAI，the Osteoarthritis Initiative）数据集这两个公开数据集上评估了该方法，结果表明，团队所提方法在膝关节软骨分割中取得了令人满意的性能。