DART: Design-Aware Real-Time Microfluidics

• DART is a design-aware and real-time hardware–software co-design paradigm for microfluidic live-cell imaging that exploits CAD blueprints to guide ROI localization and image analysis.
• It enables throughput-independent localization by aligning CAD designs with physical devices, rapidly mapping over a thousand ROIs while eliminating time-consuming manual adjustments.
• The pipeline integrates deep-learning-based marker detection and automatic segmentation, achieving sub-micron alignment and processing images in approximately 1.1 seconds.

Searching arXiv for the specified DART paper and a closely related smart-microscopy reference. DART, in microfluidic live-cell imaging, denotes the Design-Aware and Real-Time capable paradigm: a hardware–software co-design for microfluidic cultivation chips that aligns the chip’s CAD blueprint with the physical device on the microscope stage and then uses that alignment to drive region-of-interest localization, masking of microfluidic structures, and downstream image analysis in real time (Seiffarth et al., 16 Jun 2026). The paradigm addresses two bottlenecks that otherwise scale with the number of Regions of Interest (RoIs): manual or semi-automated RoI localization after chip mounting, and manual or semi-automated masking of supply channels, walls, pillars, mother-machine traps, and other microfluidic structures that are irrelevant for quantitative cell analysis. In the reported validation on the Swiss Army Knife chip, which contains eight structurally distinct RoI designs across 1164 RoI locations, DART localizes all RoIs in five minutes, removes microfluidic structures from raw microscopy images in 40 ms, and performs fully automated image analysis, including cell segmentation, in under 1.1 s per image, thereby establishing an end-to-end workflow intended for real-time-capable analysis and for closed-loop, outcome-driven smart microscopy (Seiffarth et al., 16 Jun 2026).

1. Conceptual basis and problem formulation

DART’s central idea is to exploit the CAD blueprint of the microfluidic chip during the experiment itself rather than treating each microscopy frame as a standalone image (Seiffarth et al., 16 Jun 2026). The blueprint already specifies the coordinates of all RoIs, their geometry, and the positions and shapes of interior microfluidic features. If the blueprint can be registered to the mounted chip, the stage coordinates of all RoIs can be computed from a few reference measurements, and each acquired image can be interpreted with explicit structural context.

This addresses two throughput-limiting stages in microfluidic live-cell imaging. First, chip mounting introduces an unknown offset and rotation between CAD coordinates and stage coordinates, so each experiment conventionally requires manual localization of every RoI. The paper states that this effort scales linearly with RoI count and can take hours for high-density chips. Second, raw images contain both cells and microfluidic structures, but only the cultivation area within each RoI is relevant for quantitative analysis. Even when cell segmentation is available, the segmentation model does not know where the cultivation region lies relative to walls, channels, reservoirs, or internal structures; users must therefore remove irrelevant cells and mask the microfluidics after the experiment. Together, these bottlenecks delay time-to-insight by hours to days and prevent real-time, closed-loop experiments.

DART’s response is a CAD–device–image alignment strategy. Once that alignment exists, throughput-independent RoI localization follows directly from the blueprint, and geometry-aware masking can be projected from the blueprint into the image without geometry-specific training. The paper presents this as a shift from image-only inference to design-aware analysis.

2. Hardware–software co-design and chip augmentation

DART is explicitly described as a hardware–software co-design in which the chip is made analysis-friendly at design time and the software pipeline is built to exploit that design knowledge (Seiffarth et al., 16 Jun 2026). The CAD is programmatically augmented at the CIF level so that each RoI carries a cross marker, a circle marker, and a numeric ID. The ID encodes row and column in a four-digit scheme: the first two digits are the row index and the last two digits are the column index. Marker positions and distances relative to the RoI are configurable.

The marker geometry is fixed by fabrication constraints. The cross-and-circle markers have diameter 8 μm, and the cross bar width is 2 μm; structures below 1 μm were difficult to fabricate reproducibly. These additions serve three functions stated in the paper: unique mapping between each physical RoI and the CAD via the ID; high-precision alignment of design and image coordinates through marker detection and transformation estimation; and an orientation reference via the cross–circle vector, which provides rotation and scale information.

The design is agnostic to RoI geometry. In the Swiss Army Knife chip, the CAD contains eight RoI architectures: simple box chambers, open chambers, chambers with interior pillars, mother-machine-like traps with narrow $1\,\mu\text{m}$ channels, and two-trap mother-machine variants. DART requires only the RoI polygon or polygons and the relative position of the markers; the alignment and masking procedure are otherwise unchanged across designs. This suggests that the paradigm is intended to generalize across arbitrary numbers of RoIs, arbitrary chip layouts, and heterogeneous interior structures without per-geometry retraining.

3. CAD–stage and CAD–image alignment

DART operates with two alignment levels: a coarse device-level alignment between CAD blueprint and microscope stage, and a fine image-level alignment between CAD blueprint and each acquired microscopy image (Seiffarth et al., 16 Jun 2026). Both depend on deep-learning-based fiducial detection. The marker detector is a YOLO-family model trained on 319 phase-contrast images at 100× and 0.066 μm/px, annotated with 472 cross markers and 470 circle markers, with a 60/15/25 train/validation/test split. The selected model is YOLOv26-s, 1280 px, detection task, which achieves a 95th-percentile marker-center detection error of 2.81 px = 0.18 μm and 16.4 ± 1.0 ms inference time per image on an A100 GPU.

For coarse alignment, the operator manually visits three RoIs, records one phase-contrast image and one stage coordinate per RoI, and notes the RoI ID. From the detected markers and the known CAD offsets, DART reconstructs each RoI center in image coordinates and converts it to stage coordinates. With three non-collinear point correspondences, the blueprint-to-stage mapping is an affine transformation in homogeneous coordinates: $$T(\mathbf{p}) = \mathbf{A}\cdot\mathbf{p} = \begin{bmatrix} a_{11} & a_{12} & t_1 \ a_{21} & a_{22} & t_2 \ 0 & 0 & 1 \end{bmatrix} \begin{bmatrix} x_1\ y_1\ 1 \end{bmatrix},$$ with

$$\mathbf{A} = \mathbf{X}_{\text{stage}} \cdot \mathbf{X}_{\text{blueprint}}^{-1}.$$

Once $\mathbf{A}$ is known, any RoI center in blueprint coordinates is mapped to stage coordinates by one matrix multiplication. The paper emphasizes that this step is independent of the number of RoIs.

For fine alignment, DART detects cross and circle markers in each image, enumerates cross–circle pairs, filters them by spacing consistency with the blueprint, estimates a global rotation from the cross-to-circle vectors, rotates the image, and translates the CAD polygon so that its cross-marker position coincides with the detected cross-marker location. The corresponding rigid transform is

$$\begin{bmatrix} x_{\text{img}}\ y_{\text{img}}\ 1 \end{bmatrix} = \begin{bmatrix} \cos\theta & -\sin\theta & t_x \ \sin\theta & \cos\theta & t_y \ 0 & 0 & 1 \end{bmatrix} \begin{bmatrix} x_{\text{CAD}}\ y_{\text{CAD}}\ 1 \end{bmatrix}.$$

The implementation rotates the image with Kornia or OpenCV warpAffine and rasterizes the translated blueprint polygon. The paper characterizes the result as sub-micron alignment between CAD-specified RoI structures and image content.

4. Throughput-independent localization and design-aware masking

Once the coarse affine mapping is established, all 1164 RoIs on the Swiss Army Knife chip can be localized from the blueprint in a constant-effort step requiring only the three reference measurements, which the paper reports as approximately five minutes of operator time (Seiffarth et al., 16 Jun 2026). Localization performance was evaluated by moving the stage to the predicted coordinates for every RoI, reacquiring an image, reconstructing the RoI center from marker detection, and measuring Euclidean error. The reported coarse-alignment accuracy is a median L2 localization error of 10.46 μm and a 90th percentile of 18.39 μm, against a field of view of 168.4 × 142.1 μm². The authors attribute these errors to stage inaccuracy, affine approximation, PDMS fabrication tolerances, and local chip deformation; nevertheless, the RoIs are reliably brought into the field of view for fine alignment.

Fine alignment enables blueprint-derived masking. For each image, DART rotates the image into blueprint orientation, loads the CAD polygon associated with the RoI ID, translates it so that the CAD cross-marker matches the detected cross-marker position, rasterizes the polygon into a binary mask with Shapely and Rasterio, and applies the mask to crop the image to the RoI bounding box. Microfluidic walls, pillars, mother-machine channels, and related structures are masked, while only the cultivation region remains unmasked. The paper describes this procedure as purely model-free with respect to the microfluidic geometry because the geometry is encoded in the CAD rather than learned from images.

On 1739 images, the combined runtime for marker detection, matching, rotation, masking, and cropping is 39.2 ms mean with 21.7 ms standard deviation, corresponding to approximately 25.5 FPS without segmentation. Runtime depends primarily on image size, not RoI count or chip layout. The same masking pipeline is applied to all eight Swiss Army Knife geometries with no additional training or per-design tuning. DART also includes automatic error detection by rejecting images whose detected cross–circle distances deviate excessively from the expected marker spacing.

5. Automated image analysis and quantitative outputs

On top of design-aware masking, DART implements a fully automated image-analysis pipeline that performs fine alignment, masking, cell segmentation, removal of segmentation masks overlapping microfluidic structures, and quantitative analysis (Seiffarth et al., 16 Jun 2026). Segmentation uses Cellpose-SAM on the cropped, masked RoI image. Because the masking step removes other RoIs and microfluidic background, the segmentation model operates on fewer pixels and only within the designated cultivation area.

The dominant runtime cost is segmentation. The paper reports a mean segmentation time of 1027.8 ± 695.0 ms per image and a total per-image time of 1067.0 ± 697.1 ms, giving an effective throughput of ~0.94 FPS, or approximately 1.1 s per image. Segmentation time varies with RoI size and cell count: smaller or less populated RoIs such as simple boxes are faster, while larger or highly populated geometries such as BigBox reach approximately 1.4–1.6 s. In the validation experiment, the microscope acquired one image every ~2 s, so the full DART pipeline remained faster than acquisition.

For each RoI and time point, DART computes Total Single-Cell Area (TSCA) as the sum of segmented cell-instance areas converted to $\mu\text{m}^2$. The TSCA time series is then fitted with a logistic growth model: $$N(t)=\frac{K}{1+\frac{K-N_0}{N_0}\cdot \exp(-r\cdot t)},$$ where $N_0$ is initial TSCA, $r$ is intrinsic growth rate, and $K$ is carrying capacity. Fitting uses SciPy’s curve_fit and yields $$T(\mathbf{p}) = \mathbf{A}\cdot\mathbf{p} = \begin{bmatrix} a_{11} & a_{12} & t_1 \ a_{21} & a_{22} & t_2 \ 0 & 0 & 1 \end{bmatrix} \begin{bmatrix} x_1\ y_1\ 1 \end{bmatrix},$$0, doubling time $$T(\mathbf{p}) = \mathbf{A}\cdot\mathbf{p} = \begin{bmatrix} a_{11} & a_{12} & t_1 \ a_{21} & a_{22} & t_2 \ 0 & 0 & 1 \end{bmatrix} \begin{bmatrix} x_1\ y_1\ 1 \end{bmatrix},$$1, $$T(\mathbf{p}) = \mathbf{A}\cdot\mathbf{p} = \begin{bmatrix} a_{11} & a_{12} & t_1 \ a_{21} & a_{22} & t_2 \ 0 & 0 & 1 \end{bmatrix} \begin{bmatrix} x_1\ y_1\ 1 \end{bmatrix},$$2, and $$T(\mathbf{p}) = \mathbf{A}\cdot\mathbf{p} = \begin{bmatrix} a_{11} & a_{12} & t_1 \ a_{21} & a_{22} & t_2 \ 0 & 0 & 1 \end{bmatrix} \begin{bmatrix} x_1\ y_1\ 1 \end{bmatrix},$$3. The paper explicitly notes that it does not report IoU, Dice, or precision metrics for segmentation; the quantitative evidence instead consists of qualitative segmentation results and growth curves.

6. Validation on the Swiss Army Knife chip, real-time operation, and limits

The Swiss Army Knife (SAK) chip serves as both a DART test bed and a high-throughput cultivation platform (Seiffarth et al., 16 Jun 2026). It contains four independent medium channels, each with a grid of RoIs arranged in rows and columns, and eight distinct RoI designs distributed across rows: NormalBox-inner $$T(\mathbf{p}) = \mathbf{A}\cdot\mathbf{p} = \begin{bmatrix} a_{11} & a_{12} & t_1 \ a_{21} & a_{22} & t_2 \ 0 & 0 & 1 \end{bmatrix} \begin{bmatrix} x_1\ y_1\ 1 \end{bmatrix},$$4, BigBox-inner $$T(\mathbf{p}) = \mathbf{A}\cdot\mathbf{p} = \begin{bmatrix} a_{11} & a_{12} & t_1 \ a_{21} & a_{22} & t_2 \ 0 & 0 & 1 \end{bmatrix} \begin{bmatrix} x_1\ y_1\ 1 \end{bmatrix},$$5, OpenBox-inner $$T(\mathbf{p}) = \mathbf{A}\cdot\mathbf{p} = \begin{bmatrix} a_{11} & a_{12} & t_1 \ a_{21} & a_{22} & t_2 \ 0 & 0 & 1 \end{bmatrix} \begin{bmatrix} x_1\ y_1\ 1 \end{bmatrix},$$6, Mothermachine-1x-inner $$T(\mathbf{p}) = \mathbf{A}\cdot\mathbf{p} = \begin{bmatrix} a_{11} & a_{12} & t_1 \ a_{21} & a_{22} & t_2 \ 0 & 0 & 1 \end{bmatrix} \begin{bmatrix} x_1\ y_1\ 1 \end{bmatrix},$$7, NormalBox-pillar-inner $$T(\mathbf{p}) = \mathbf{A}\cdot\mathbf{p} = \begin{bmatrix} a_{11} & a_{12} & t_1 \ a_{21} & a_{22} & t_2 \ 0 & 0 & 1 \end{bmatrix} \begin{bmatrix} x_1\ y_1\ 1 \end{bmatrix},$$8 with interior pillars, BigBox-pillar-inner $$T(\mathbf{p}) = \mathbf{A}\cdot\mathbf{p} = \begin{bmatrix} a_{11} & a_{12} & t_1 \ a_{21} & a_{22} & t_2 \ 0 & 0 & 1 \end{bmatrix} \begin{bmatrix} x_1\ y_1\ 1 \end{bmatrix},$$9 with pillars, OpenBox-collector-inner $$\mathbf{A} = \mathbf{X}_{\text{stage}} \cdot \mathbf{X}_{\text{blueprint}}^{-1}.$$0 with collector features, and Mothermachine-2x-inner $$\mathbf{A} = \mathbf{X}_{\text{stage}} \cdot \mathbf{X}_{\text{blueprint}}^{-1}.$$1. The same DART pipeline is used across all eight.

In the live-cell experiment with Corynebacterium glutamicum, DART processed 1739 images in 31 minutes and segmented >500,000 individual cells. Relative to a total imaging time of 540 minutes, the paper reports a real-time factor of 17.4. This runtime profile underlies the claim that DART is compatible with online quantitative readouts and with closed-loop or outcome-driven experiments in structured microfluidic systems.

The paper also states several limitations. Coarse stage-level localization has ~10–18 μm median and 90th-percentile error, which can be problematic if the RoI fills most of the field of view. Very fine structures, especially mother-machine features around 1 μm width, remain sensitive to small alignment errors; even the reported 0.18 μm marker-detection error can slightly occlude or expose narrow traps. The method assumes that the fabricated chip matches the CAD geometry up to an affine transform plus modest local deformation, so PDMS deformation and fabrication defects can degrade mapping quality. Marker visibility is required, and poor focus, low contrast, or defective markers can cause detection failures, although spacing-based error checks mitigate incorrect masks. Finally, segmentation remains the runtime bottleneck, and the paper does not establish segmentation accuracy across different cell types or imaging modalities.

The stated future directions follow directly from these constraints: additional local markers around each RoI for more robust local alignment; faster or task-specific segmentation models; more reference RoIs or continuous image-based corrections for improved stage mapping; extension to other structured live-cell imaging platforms such as organ-on-chip devices or multiwell plates with structured microfluidics; and integration with closed-loop control frameworks for fully autonomous experiments. A plausible implication is that DART’s primary contribution is not a single algorithmic module but a shift in experimental design: the microfluidic device is fabricated so that structural prior information remains available to the analysis stack throughout acquisition.