🤖 AI Summary
To address the limited specificity and throughput of label-free cell classification in microfluidic systems, this paper proposes a multiplexed image-based machine learning framework targeting cellular mechanophenotypic heterogeneity. The method innovatively integrates microfluidic deformation imaging with multiscale convolutional feature extraction, enabling, for the first time, decoupled modeling of mechanical microenvironment response images and multiplexed mechano-morphological features. It further incorporates a cross-modal attention mechanism and a weakly supervised multiple-instance learning (MIML) paradigm to achieve label-free, non-invasive single-cell subtype identification. Evaluated across multiple cancer cell lines, the framework achieves 98.2% classification accuracy—substantially outperforming conventional morphological or single-mechanical-parameter approaches. This work establishes a new paradigm for high-purity cell sorting and functional phenotyping, advancing label-free mechanobiological analysis in microfluidics.
📝 Abstract
This paper has been withdrawn by Khayrul Islam.