Scholarlink

The structural maturation of sarcomeres in human induced pluripotent stem cell-derived cardiomyocytes (hiPSC-CMs) is a critical determinant of their functional viability for disease modeling, drug screening, and regenerative therapy. However, current models face challenges, a limited ability to simultaneously capture both global and fine-grained sarcomeric structures, high sensitivity to image quality variations and morphological irregularities, and inadequate utilization of biological metadata and lack of integration of biological context. To overcome these limitations, we propose SAMBA-Net (Sarcomere Analysis via Multi-modal Bi-Mamba Attention fusion Network), an image-text based deep learning framework that combines multi-scale visual features and structured biological context. Specifically, SAMBA-Net incorporates: a Structure-Guided Module that extracts global image features enhanced by low-frequency components from the Stationary Wavelet Transform(SWT) and periodic patterns from the Fast Fourier Transform, a Texture-Fused Module that captures localized patterns using Gabor filters, Laplacian of Gaussian filtering, and high-frequency decompositions of SWT, and a Context-Enriched Module that reformulates biological metadata into natural language sentences and allows the model to effectively leverage semantic relationships within the biological context. The multi-modal representations are fused via a novel Mamba-Attention Fusion3 Module, enabling the model to learn both shared and modality-specific patterns. SAMBA-Net outputs a continuous sarcomere organization score ranging from 1.0 to 5.0, reflecting the structural maturity of each cardiomyocyte. Experimental results on the Allen Institute hiPSC-CM dataset show that SAMBA-Net achieves superior performance, with a Spearman correlation of 0.870, MAE of 0.259, MSE of 0.119, and R2 of 0.749, outperforming existing models such as SarcNet and D-SarcNet. These findings demonstrate the effectiveness of modality-aware textual reformulation and deep fusion for robust, interpretable, and fine-grained assessment of sarcomere organization in hiPSC-CMs.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

SAMBA-Net: Enhancing Sarcomere Organization Evaluation with AI-Powered Multi-Modal Learning

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Yingxin Chen,
  • Hongpeng Yang,
  • Shiqiang Ma,
  • Fei Guo

摘要

The structural maturation of sarcomeres in human induced pluripotent stem cell-derived cardiomyocytes (hiPSC-CMs) is a critical determinant of their functional viability for disease modeling, drug screening, and regenerative therapy. However, current models face challenges, a limited ability to simultaneously capture both global and fine-grained sarcomeric structures, high sensitivity to image quality variations and morphological irregularities, and inadequate utilization of biological metadata and lack of integration of biological context. To overcome these limitations, we propose SAMBA-Net (Sarcomere Analysis via Multi-modal Bi-Mamba Attention fusion Network), an image-text based deep learning framework that combines multi-scale visual features and structured biological context. Specifically, SAMBA-Net incorporates: a Structure-Guided Module that extracts global image features enhanced by low-frequency components from the Stationary Wavelet Transform(SWT) and periodic patterns from the Fast Fourier Transform, a Texture-Fused Module that captures localized patterns using Gabor filters, Laplacian of Gaussian filtering, and high-frequency decompositions of SWT, and a Context-Enriched Module that reformulates biological metadata into natural language sentences and allows the model to effectively leverage semantic relationships within the biological context. The multi-modal representations are fused via a novel Mamba-Attention Fusion3 Module, enabling the model to learn both shared and modality-specific patterns. SAMBA-Net outputs a continuous sarcomere organization score ranging from 1.0 to 5.0, reflecting the structural maturity of each cardiomyocyte. Experimental results on the Allen Institute hiPSC-CM dataset show that SAMBA-Net achieves superior performance, with a Spearman correlation of 0.870, MAE of 0.259, MSE of 0.119, and R2 of 0.749, outperforming existing models such as SarcNet and D-SarcNet. These findings demonstrate the effectiveness of modality-aware textual reformulation and deep fusion for robust, interpretable, and fine-grained assessment of sarcomere organization in hiPSC-CMs.