AI-CMCA: a deep learning-based segmentation framework for capillary microfluidic chip analysis

Abstract

Capillary microfluidic chips (CMCs) enable passive liquid transport via surface tension and wettability gradients, making them central to point-of-care diagnostics and biomedical sensing. However, accurate analysis of capillary-driven flow experiments remains constrained by the labour-intensive, time-consuming, and inconsistent nature of manual fluid path tracking. Here, we present AI-CMCA, an artificial intelligence framework designed for capillary microfluidic chip analysis, which automates fluid path detection and tracking using deep learning-based segmentation. AI-CMCA combines transfer learning-based feature initialization, encoder-decoder-based semantic segmentation to recognize fluid in each frame, and sequential frame analysis to track then quantify fluid progression. Among the five tested architectures, including U-Net, PAN, FPN, PSP-Net, and DeepLabV3+, the U-Net model with MobileNetV2 achieved the highest performance, with a validation IoU of 99.24% and an F1-score of 99.56%. Its lightweight design makes it well suited for smartphone or edge deployment. AI-CMCA demonstrated a strong correlation with manually extracted data while offering superior robustness and consistency in fluid path analysis. AI-CMCA performed fluid path analysis up to 100 times faster and over 10 times more consistently than manual tracking, reducing analysis time from days to minutes while maintaining high precision and reproducibility across diverse CMC architectures. By eliminating the need for manual annotation, AI-CMCA significantly enhances efficiency, precision, and automation in microfluidic research.

Keywords: Capillary microfluidic chip; Deep learning; Fluid path detection; Image segmentation; Passive fluid transport; Point-of-care diagnostics.

Scheme 1

Illustrating (A) a representative experimental setup featuring a set of parallel straight channels as a fundamental microfluidic element in a capillary microfluidic chip (CMC) design, (B) AI-CMCA framework and the sequential steps, including (a) video acquisition, (b) preprocessing, (c) AI model implementation, and (d) fluid path analysis, determining the fluid path length variation over time in a capillary-driven fluid flow over CMC.

Fig. 1

Schematic of the simple CMC designs and their fabrication process. (A) Design of the capillary channel network and fundamental components of straight and serpentine channels, including merging and splitting. (B) Laser cutting of individual layers. (C) Multi-layer assembly of CMC via pressure-sensitive adhesive sheets. (D) Experimental setup for recording the operation of the assembled CMC.

Fig. 2

Data processing and segmentation model evaluation. (A) Workflow for preparation and processing data from raw video frames. (B) Comparison of segmentation models and encoder architectures with associated performance metrics. (C) Performance trends across training epochs for IoU, F1-score, and Dice loss. (D) Representative inputs, ground-truth masks, predicted outputs, and heatmaps. Scale bar: 1 cm.

Fig. 3

Workflow for fluid segmentation and front tracking in CMCs (A) (a) Data pre-processing: a1: applying noise reduction to raw videos, a2: extracting individual frames. (b) Data processing: applying the trained U-Net model to segment fluid regions from the chip background. (c) Data post-processing: c1: frame differencing to isolate newly advanced fluid areas and clustering to remove extraneous pixels, c2: generating a heatmap of detected flow regions and tracking center points (red dots) over time, with an inset plot illustrating fluid progression. (B) Pseudocode illustrating the fluid front tracking algorithm, including clustering, region labelling, and center-point computation for each labelled region.

Fig. 4

Automated flow path analysis and efficiency gains in different CMC designs. (A) Quantitative flow analysis of a CMC with six parallel channels: (a) results for br0 and (b) results for br1, comparing system-generated and user-tracked data. (B) Flow tracking in a CMC with splitting and merging regions, showing system- and user-generated results across four branches. (C) Flow tracking in a tree-shaped CMC featuring ten merging points and fifteen splitting points, representing a real-life CMC with a complex channel network: (a) extracted datapoints and (b) a photograph of the CMC highlighting critical points in red circles and branch numbers in green. (D) Comparison of fluid front tracking accuracy between AI-CMCA-generated and manually selected user-annotated center points for both (a) simple straight-channel CMC and (b) more complex split-merge-serpentine CMC. (E) Total time required for analysis across different CMC designs such as (a) straight channels, (b) split/merge, (c) serpentine channels, (d) split merge serpentine, SARS-Cov-2-POC testingand GFAP POC testinghighlighting the efficiency gains of the automated system over manually tracking.