Source-linked AI summary
Data-Efficient and Interpretable Classification of Circulating Tumor Cell Phenotypes in Microfluidic Devices via Deep Learning
Serena Su, Yifan Wang, Senwei Liang
TL;DR
Classifying CTC phenotypes from microfluidic trajectories is difficult because nonlinear dynamics obscure biophysical differences and available data are limited. The paper introduces SubSeq augmentation with Grad-CAM interpretation, finding improved classification and informative localized trajectory segments.
Problem
Nonlinear, coupled trajectory dynamics and limited data make conventional analytical or threshold-based CTC phenotype classification difficult.
Method
The framework trains a classifier on randomly sampled contiguous trajectory segments and uses Grad-CAM to map predictions to trajectory features and device regions.
Results
SubSeq consistently improves classification across training-data regimes, while localized segments, especially velocity features, provide the strongest predictive information.
Takeaways & Limitations
Localized cell–flow–structure interactions contain sufficient biophysical information for classification, with full trajectories showing substantial redundancy.
Takeaways & Limitations
The simulated dataset is small, represents only two mechanical phenotypes, and leaves generalization to experimental cells and broader properties uncertain.
Abstract
from arXiv · showhide
Accurate classification of circulating tumor cell (CTC) phenotypes can provide valuable information for assessing metastatic potential. Label free microfluidic devices provide a hydrodynamic obstacle course that transforms subtle biophysical characteristics of CTCs, including size and deformability, into distinct kinematic trajectories. However, the highly nonlinear fluid structure interactions governing these trajectories make the inverse problem of inferring cellular phenotype from trajectory data analytically intractable. While deep neural networks (DNNs) have emerged as a powerful approach for addressing this inverse problem, their effectiveness is constrained by the limited availability of trajectory data and the lack of physical interpretability. To address these challenges, we propose an interpretable and data efficient DNN framework for trajectory based CTC classification. To mitigate the scarcity of data, we develop Subsequence (SubSeq), a targeted augmentation strategy that randomly extracts informative local trajectory segments during training to promote learning from localized patterns. We further apply Gradient Weighted Class Activation Mapping to identify the trajectory features and physical regions of the microfluidic device that drive model predictions. Experimental results demonstrate that SubSeq improves classification accuracy over the evaluated baseline and augmentation methods. Furthermore, interpretability analysis suggests that localized trajectory segments contain substantial biophysical information relevant to accurate classification. This provides justification for SubSeq and also highlights the redundancy of full-length trajectories. More broadly, the proposed framework views microfluidic geometries as physical encoders of cellular mechanical properties, providing mechanistic insights that may inform the future design of diagnostic devices.
1. Introduction
CTC trajectories encode phenotype-relevant mechanical properties, but overlapping properties produce nonlinear dynamics that challenge conventional classification. The proposed SubSeq and Grad-CAM framework addresses limited data and black-box interpretation while identifying informative trajectory signals and localized interactions.
- Motivation: CTCs are important mediators of metastasis, and their detection and classification support cancer diagnosis, prognosis, treatment response, and recurrence-risk assessment.The paper situates CTC analysis within liquid-biopsy technology development.
- Challenge: Overlapping biophysical properties across phenotypes generate highly nonlinear, coupled trajectory dynamics that resist conventional analytical models and threshold-based sorting.The resulting motion patterns are difficult to describe or classify with simple strategies.
- Limitations: The approach targets two limitations: computationally expensive simulations restrict trajectory-data availability, while DNNs remain difficult to interpret as black boxes.These limitations may constrain model generalization and transparency.
- Contributions: The framework introduces Subsequence (SubSeq) sampling, which randomly extracts continuous trajectory segments during training, and integrates Grad-CAM for interpretable prediction analysis.SubSeq is designed to alleviate scarce simulation data, while Grad-CAM maps trajectory segments to model predictions.
- Key findings: Cell velocity carries primary predictive signals, position provides complementary information, and localized cell–flow–structure interactions contain sufficient biophysical information for phenotype classification.Combining velocity and position yields the highest classification performance, while full trajectories show significant redundancy.
2. Related Work
Related work frames label-free microfluidics as platforms for both rare-cell detection and phenotypic characterization, while emphasizing specialized augmentation for time-series data and interpretable DNN evaluation. These foundations motivate trajectory-based classification methods that account for physical and temporal structure.
- Microfluidic phenotyping: Simulation frameworks model how flow conditions, device geometry, ligand density, and cell properties govern CTC capture dynamics.Tan et al. integrated physical modeling with device design to study CTC transport and adhesion.
- Microfluidic phenotyping: Microfluidic platforms increasingly characterize clinically relevant phenotypes beyond simple cell isolation, including size, deformability, adhesion, metabolism, and surface markers.This review describes the field’s shift toward extracting biologically meaningful cellular information.
- Data augmentation: Data augmentation expands training diversity and can mitigate overfitting in data-scarce environments by exposing models to broader input variations.The passage contrasts conventional computer-vision transformations with the broader goal of improving generalization.
- Data augmentation: Time-series augmentation must preserve strict temporal dependencies because arbitrary transformations can distort sequential dynamics and produce invalid samples.Specialized approaches include time-domain, frequency-domain, and learning-based methods; time-domain methods manipulate raw sequences directly.
- Model interpretability: Interpreting DNN decisions is important for model evaluation in medical diagnosis, with Class Activation Maps highlighting image regions contributing to target-class predictions.The original CAM method generates class-specific activation maps by linearly combining learned features.
3. Preliminary
This section defines trajectory-based classification of soft and hard circulating tumor cell phenotypes and describes the CNN testbed used to evaluate augmentation and interpretability. Trajectories are generated through computationally demanding fluid–structure interaction simulations, limiting available training data.
- Data generation and motivation: Computationally demanding fluid–structure interaction simulations produce limited training samples, motivating effective augmentation strategies to improve DNN generalization.The simulations couple incompressible Newtonian blood plasma dynamics with deformable CTC membrane models.
- Classification task: The task maps each cell trajectory to a phenotype label distinguishing soft and hard CTC types.Each trajectory contains position and velocity time series, while labels p^(i) ∈ {0, 1} denote the two phenotypes.
- CNN framework: The adopted CNN processes position and velocity sequences through parallel branches, concatenates extracted features, and classifies phenotypes with a multilayer perceptron.The architecture serves as the primary testbed for comparing data augmentation with baseline training and applying Grad-CAM.
- CNN framework: Each CNN branch applies one-dimensional convolutions, ReLU activations, max pooling, and adaptive average pooling to its three-channel trajectory sequence.The first convolution uses 256 filters with kernel size 3, followed by max pooling with kernel size 2; a second convolution uses 256 filters and kernel size 3.
4. Proposed Methods for Improving and Interpreting CTC Classification
This section presents SubSeq, which augments training by exposing classifiers to randomly selected trajectory segments, and Grad-CAM, which identifies influential temporal features and maps their importance into physical device space.
- SubSeq: SubSeq randomly extracts contiguous temporal segments so classifiers learn discriminative local motion patterns instead of relying on complete trajectories.The method limits trajectory information during training while preserving a segment-based input.
- SubSeq: The subsequence length is uniformly sampled from L_min through T, with its minimum set by L_min = max(⌊rT⌋, 1), and the start index sampled uniformly.The minimum subsequence ratio r controls the amount of information retained in each sample.
- SubSeq: Extracted subsequences are zero-padded to length T, allowing training on partial trajectories while maintaining a consistent CNN input size.Only a randomly selected portion of each trajectory is used by the classifier.
- Grad-CAM: Grad-CAM uses backpropagated gradients and intermediate feature maps to localize trajectory regions that contribute most to each model prediction.The CNN’s separate position and velocity branches are analyzed independently.
- Grad-CAM: Temporal importance scores from both branches are averaged within spatial grid boxes to produce a physical-space heatmap of trajectory importance.Each time-step score is associated with its corresponding position or velocity before spatial aggregation across test trajectories.
5. Experiments
Experiments show that SubSeq improves classification performance and stability, especially when training data are scarce, while velocity provides the core predictive signal and position adds complementary information. Grad-CAM further indicates that predictions rely on localized, phenotype-specific trajectory patterns and channel regions rather than entire trajectories uniformly.
- SubSeq across diverse data regimes: SubSeq consistently outperforms the baseline across train-test regimes, with higher median testing accuracy and ROC-AUC, narrower distributions, and a larger advantage when training data are limited.The narrower performance spread indicates greater stability across independent trials.
- SubSeq against alternative data augmentation methods: Compared with Cutout and Mixup, SubSeq achieves competitive central performance while maintaining reduced variability and mitigating the worst-case performance observed with Cutout.For testing accuracy, SubSeq has a comparable median to Cutout, tighter distribution, and higher lower bound; for ROC-AUC, Cutout has the highest peak and median but largest variability.
- Effect of input features: position vs. velocity: Velocity-only inputs remain highly competitive, whereas position-only inputs perform significantly worse; combining position and velocity consistently yields the strongest overall predictive capability.The unaugmented combined-input baseline has ROC-AUC 0.8612, augmentation raises velocity-only AUC to 0.9076, and the augmented full model reaches a peak AUC of 0.9347.
- Grad-CAM interpretation: Trajectory-level Grad-CAM assigns high importance to selected portions rather than entire trajectories, indicating that localized trajectory patterns drive classification predictions.This supports the interpretation that full-length trajectories are not uniformly informative.
- Grad-CAM interpretation: Spatial Grad-CAM maps show that phenotype-discriminative information is concentrated in localized channel regions, with velocity producing more localized importance than position and soft and hard cells exhibiting different high-importance regions.The findings suggest that the model learns phenotype-specific signatures based on local motion responses to the surrounding fluid and microposts.
6. Discussion & Conclusion
The study presents an interpretable, data-efficient framework for trajectory-based CTC phenotype classification that addresses limited simulated data and opaque neural-network predictions. SubSeq improves generalization by emphasizing localized trajectory information, while physical-domain interpretability offers mechanistic insight into microfluidic encoding and motivates future validation.
- Framework: The framework addresses limited simulated training data and the lack of transparency in deep neural network predictions for trajectory-based CTC classification.It is designed for phenotype classification in a hyperuniform microfluidic device.
- SubSeq: SubSeq consistently improves classification performance across training-data regimes and provides stable, reliable generalization compared with existing augmentation methods.The results support SubSeq as a data-efficient strategy under data-scarce conditions.
- Interpretability: Interpretability analysis indicates that accurate predictions rely primarily on localized trajectory segments rather than complete trajectories.Grad-CAM visualizations connect prediction-relevant trajectory information to physical regions of the device.
- Physical interpretation: The hyperuniform micropost array acts as a physical encoder that transforms intrinsic cellular mechanical properties into trajectory signatures decoded by machine-learning models.Mapping learned importance to the physical domain provides mechanistic insight into where and how discriminative information arises.
- Limitations and future work: The study is limited by a small, simulation-only dataset containing trajectories for only two mechanical phenotypes.Future work includes experimental trajectory data, more diverse cellular characteristics and geometries, evaluation across neural-network architectures, and incorporating interpretability into device-architecture optimization.
- Broader significance: The combination of data-efficient learning and physically grounded interpretability is presented as a general framework for trajectory-based analysis in microfluidics and other scientific applications.The stated objective is to support both predictive performance and mechanistic understanding.
Appendix
The appendix describes method-specific tuning of training parameters and auxiliary strategies, with SubSeq, Cutout, and Mixup evaluated under specified configurations. SubSeq uses a high activation probability and extracts sequences no shorter than 60% of the full trajectory.
- Training configuration: Each augmentation method was evaluated under its own best-performing configuration, with scheduling and noise injection applied selectively when they improved fine-tuning performance.Hyperparameters and auxiliary strategies were tuned independently for each method.
- SubSeq configuration: SubSeq used a learning rate of 3×10−4, an activation probability of 0.99, and a minimum length ratio of 0.6.The minimum length ratio is measured relative to the full sequence length.
- Alternative augmentations: Under a fixed 80∕20 train-test split, Cutout used a learning rate of 3 × 10−4, a masking probability of 0.99, and a maximum mask width of 250.These settings were used for comparison against SubSeq.
- Alternative augmentations: Mixup used a learning rate of 5 × 10−4 and a mixing coefficient of 𝛼= 0.5.The Mixup configuration was evaluated in the same fixed 80∕20 train-test comparison.
CRediT authorship contribution statement
The authors contributed across conceptualization, methodology, software, data curation, and original draft preparation. Yifan Wang additionally handled data curation.
- Serena Su contributed methodology, software, and original draft preparation.
- Yifan Wang contributed study conceptualization, methodology, data curation, and original draft preparation.
- Senwei Liang contributed study conceptualization, methodology, software, and original draft preparation.