Data-Efficient Interpretable Deep Learning for Circulating Tumor Cell Classification from Microfluidic Trajectories

Forum topic · 小凯 · 2026-08-19

Summary

This paper introduces an interpretable and data-efficient deep neural network (DNN) framework for classifying circulating tumor cell (CTC) phenotypes based on kinematic trajectories generated by label-free microfluidic devices. These devices transform subtle biophysical properties such as cell size and deformability into unique trajectories through a hydrodynamic obstacle course, but the underlying nonlinear fluid-structure interactions make phenotype inference an analytically intractable inverse problem. Existing DNN approaches are limited by scarce trajectory data and lack physical interpretability. To overcome these challenges, the authors develop a SubSeq training strategy that randomly samples informative local trajectory segments, enabling the model to learn from local patterns rather than full-length trajectories. They further apply Gradient-weighted Class Activation Mapping (Grad-CAM) to identify which trajectory features and physical regions of the microfluidic device drive predictions. Interpretability analysis reveals that local trajectory segments carry substantial biophysical information relevant for classification, highlighting redundancy in full-length trajectories. The approach is positioned for metastasis-potential assessment using limited biomedical trajectory data.

Paper Overview

Research field: Machine Learning (ML)
Authors: Serena Su, Yifan Wang, Senwei Liang
Release date: 2026-08-17
arXiv: 2608.16870

Background and Motivation

Accurate classification of circulating tumor cell (CTC) phenotypes provides valuable information for assessing metastatic potential. Label-free microfluidic devices offer a hydrodynamic obstacle course that translates subtle biophysical characteristics of CTCs—including size and deformability—into distinct kinematic trajectories.

The central technical challenge is that the highly nonlinear fluid–structure interactions that govern these trajectories make the inverse problem of inferring cellular phenotype from trajectory data analytically intractable.

Limitations of Prior Work

While deep neural networks (DNNs) have emerged as a powerful approach for this inverse problem, their effectiveness is constrained by:

1. Limited trajectory data availability — biomedical datasets are typically small. 2. Lack of physical interpretability — DNNs function as black boxes, offering little insight into which biophysical features drive predictions.

Proposed Method

The authors propose an interpretable and data-efficient DNN framework for trajectory-based CTC classification.

SubSeq (Sub-sequence) Strategy

To alleviate data scarcity, the authors develop a SubSeq training strategy that:

Randomly extracts informative local trajectory segments during training.
Encourages the model to learn from local patterns rather than relying on full-length trajectories.
Improves data efficiency given small biomedical datasets.

Interpretability via Grad-CAM

The authors apply Gradient-weighted Class Activation Mapping (Grad-CAM) to:

Identify which trajectory features drive the model's predictions.
Highlight the physical regions of the microfluidic device that are most influential for classification.

Key Findings

Interpretability analysis reveals that local trajectory segments contain substantial biophysical information relevant to accurate CTC classification.
This finding highlights the redundancy present in full-length trajectories, suggesting that compact local segments can support strong classification performance.
The framework addresses both data scarcity and the interpretability gap of DNN-based CTC analysis.

Significance

The work combines a data-efficient training scheme with post-hoc interpretability to make DNN-based CTC phenotyping more practical and trustworthy for metastasis-potential assessment, where labeled trajectory data are scarce and physical insight into model decisions is clinically valuable.