目的 海马体积很小，对比度极低，传统标记融合方法选用手工设计的特征模型，难以提取出适应性好、判别性强的特征。近年来，深度学习方法取得了极大成功，基于深度网络的方法已应用于医学图像分割中，但海马结构复杂，子区较多且体积差别较大，特别是CA2和CA3子区体积极小，常见的深度网络无法准确分割海马子区。为了解决这些问题，提出一种结合多尺度输入和串行处理神经网络的海马子区分割方法。方法 针对海马中体积差距较大的子区，设计两种不同的网络，结合多种尺度图像块信息，为小子区建立类别数量均衡的训练集，避免网络被极端化训练，最后，采用串行标记的方式对海马子区进行分割。结果 在Tail，SUB和PHG子区上的准确率达到了0.865，0.81，0.773，较现有的多图谱子区分割方法有较大提高，并且将体积较小子区CA2，CA3上的准确率分别提高了6%和9%。结论 该算法将基于卷积神经网络的分类方法引入到标记融合阶段，根据海马子区特殊的灰度及结构特点，设计两种针对性网络，实验证明，该算法能提取出适应性好、判别性强的特征，提高了分割准确率。

Objective Hippocampus is a structure in the brain in memory consolidation and can be divided into nine subfields. Hippocampus atrophy has been mostly studied in various neurological diseases, such as Alzheimer's disease and mild cognitive impairment. Accurate hippocampus subfield segmentation in magnetic resonance (MR) images plays a crucial role in the diagnosis, prevention, and treatment of neurological diseases. However, the segmentation is a challenging task due to small size, relatively low contrast, complex shape, and indistinct boundaries of hippocampus subfields. Numerous scholars have been engaged in hippocampus subfield segmentation. Multi-atlas-based methods can obtain accurate segmentation results by fusing propagated labels of multiple atlases in a target image space. However, the performance of multi-atlas significantly relies on the effectiveness of the label fusion method. Deep learning algorithms have emerged as promising machine-learning tools in general imaging and computer vision domains, such as medical image segmentation. However, Cornu Ammonis (CA)2 and CA3 are limited by MR resolution and are thus significantly smaller than other subfields in hippocampus MR images. Most deep learning algorithms with identical network models and uniform patches present poor segmentation accuracy regardless of the considerable differences in the volume of different subfields. Method This study proposes a combined multi-scale patch and cascaded convolutional neural network (CNN)-based classification algorithm for segmenting the hippocampus into nine subfields to address the aforementioned deficiencies. In comparison with traditional label fusion method, the proposed method does not rely on explicit features but learns to extract important features for classification. Two different CNNs are designed considering the significant volume differences among different subfields. Network 1, which considers large patches as inputs, is trained to segment large subfields accurately. Network 2, which includes two patch types with small sizes that form a two-pathway network, is trained to obtain high segmentation accuracy on small subfields. Each network is trained using datasets from multiple atlases. The same number of patches is randomly extracted from different subfields that comprise a balanced training set to manage imbalanced datasets, in which the training patches of CA2 and CA3 are less than those of other subfields. The segmentation is performed slice by slice along the coronal direction, and a two-phase cascaded segmentation procedure is designed. First, the preliminary segmentation is performed using Network 1. Second, the voxels in small subfields are further classified using Network 2. The prior structural knowledge is used to recognize and correct mislabeling voxels to further improve the segmentation accuracy. Result All experiments are validated on CIND dataset, which contains 32 subjects with manually labeled ROIs. After the preprocessing, three separate training sets, which consist of patches with three different sizes, are extracted from multiple atlases. The training data of different classes are numerically balanced. We investigate the influence of using different sizes of input patches and kernels at the first convolutional layer of Network 1 to select the most appropriate parameters. We evaluate the different segmentation performances on small subfields by using the two different networks. The Dice similarity coefficient is selected as the evaluation metric. Several approaches in recent literature are implemented to compare with the proposed algorithm. Quantitative and qualitative comparisons demonstrate that the proposed method outperforms the traditional label fusion method on most subfields. The accuracies on tail, SUB, and PHG subfields are 0.865, 0.81, and 0.773, respectively. On the two small subfields, CA2 and CA3, the proposed method exhibits better performances with accuracies of 60% and 64%, surpassing the traditional method by 6% and 9%, respectively. Conclusion In this paper, we propose a combined multi-scale patch and cascaded CNN-based method for hippocampus subfield segmentation. Two different networks are developed to solve the problems related to significant volume differences among different subfields. Three different sizes are used as inputs of two networks to capture considerable contextual information. A balanced training set is established to avoid incorrect training of the networks. We describe a cascaded segmentation procedure using the proposed networks. This cascaded procedure integrates two different networks for accurate segmentation on different subfields. The experiments show that the two CNNs with nonuniform patches outperform an identical network with a uniform patch. The significant improvement compared with traditional method shows that the feature extracted by the proposed network is considerably more effective and distinctive. Therefore, the proposed algorithm can label with high accuracy and is highly appropriate for hippocampus subfield segmentation.