MLN-net: A multi-source medical image segmentation method for clustered microcalcifications using multiple layer normalization

TL;DR

MLN-net tackles domain shift in clustered microcalcification segmentation by combining a Bézier-curve based source-domain augmentation and grayscale-inversion to generate multi-domain data, with a segmentation network that uses multiple LN layers to encode domain information. A branch-selection strategy based on cosine similarity chooses the best LN-branch for a given target domain, enabling cross-domain generalization without labeled target data. On private FFDM-DBT and public CBIS-DDSM datasets, MLN-net achieves state-of-the-art segmentation metrics, outperforming basic, specialized, and DG baselines. The approach offers a practical pathway toward robust, cross-center breast cancer screening with minimal retraining.

Abstract

Accurate segmentation of clustered microcalcifications in mammography is crucial for the diagnosis and treatment of breast cancer. Despite exhibiting expert-level accuracy, recent deep learning advancements in medical image segmentation provide insufficient contribution to practical applications, due to the domain shift resulting from differences in patient postures, individual gland density, and imaging modalities of mammography etc. In this paper, a novel framework named MLN-net, which can accurately segment multi-source images using only single source images, is proposed for clustered microcalcification segmentation. We first propose a source domain image augmentation method to generate multi-source images, leading to improved generalization. And a structure of multiple layer normalization (LN) layers is used to construct the segmentation network, which can be found efficient for clustered microcalcification segmentation in different domains. Additionally, a branch selection strategy is designed for measuring the similarity of the source domain data and the target domain data. To validate the proposed MLN-net, extensive analyses including ablation experiments are performed, comparison of 12 baseline methods. Extensive experiments validate the effectiveness of MLN-net in segmenting clustered microcalcifications from different domains and the its segmentation accuracy surpasses state-of-the-art methods. Code will be available at https://github.com/yezanting/MLN-NET-VERSON1.

Figures (9)

• Figure 1: Examples of images from FFDM and DBT imaging modalities. The images are from the same patient and represent the same area of breast tissue. Clustered microcalcifications can be seen in both imaging modalities. Compared to FFDM, DBT performs slice photography of the lesion with a much narrower interval of 1mm, producing dozens to hundreds of images depending on individual differences.
• Figure 2: The schematic diagram of MLN-net framework. MLN-net utilizes source domain data for generalizable segmentation on unseen domain data. MLN-net is composed of a source domain data augmentation method, a segmentation network, and a branch selection strategy. During training stage, the source domain data augmentation method is utilized to augment the source domain data. The augmented data is then fed to the segmentation network with multiple LN layers, and the loss functions is used to optimize the segmentation results of different branches. During testing stage, the test data is fed into the trained segmentation network. And the branch selection strategy is adopted to choose the optimal segmentation results.
• Figure 3: Examples of source-similar and source-dissimilar transformations results on FFDM-DBT dataset. In the first row, the red line signifies the transformation of source-similar based on the Bézier curve. This curve has four critical points; P1 denotes the starting point while P4 denotes the endpoint. P2 and P3 denote the control points that determine the curvature of the Bézier curve. In contrast, the blue line signifies transformation of source-dissimilar based on grayscale-inversion. And the second and the third row are the images after the transformations
• Figure 4: The composition structure of the backbone network and the Swinunet blocks in the Swinunet model.
• Figure 5: Clustered microcalcifications segmentation results of different methods. In the input and ground truth images, the red markings indicate the lesions annotated by doctors. The rest of the red markings represent the lesions identified by the different methods.