BioLite U-Net Architecture for Embedded Bioprinting

• BioLite U-Net is a lightweight encoder–decoder network using depthwise separable convolutions for efficient, real-time segmentation in bioprinting applications.
• It employs a shallow U-Net topology with two downsampling blocks and skip-connections to accurately demarcate nozzle, bioink, and background elements.
• The architecture achieves competitive segmentation metrics (mIoU ≈93%, Dice ≈96%) while reducing parameters by over 1300× compared to MobileNet-based variants.

BioLite U-Net is a lightweight, encoder–decoder convolutional architecture specifically crafted for real-time semantic segmentation on resource-constrained, embedded devices in in situ bioprinting applications. Its design optimizes parameter and compute efficiency using depthwise separable convolutions and a shallow network topology, while preserving segmentation accuracy sufficient for critical tasks such as nozzle, bioink, and background demarcation during the bioprinting process (Haider et al., 8 Sep 2025).

1. Architectural Fundamentals

BioLite U-Net adopts the canonical U-Net topology: an encoder path that progressively downsamples and abstracts image features via convolutional layers and max-pooling, coupled with a decoder path that symmetrically upsamples the feature map resolution, restoring spatial detail. It maintains skip-connections at each encoder–decoder level, crucial for transferring high-resolution spatial context needed for accurate segmentation of small and thin structures such as the nozzle tip or bioink filaments.

Crucially, every convolutional layer is replaced by a depthwise separable convolution, decomposing a standard conv block into (i) a depthwise convolution applied independently to each channel for local spatial filtering, and (ii) a pointwise 1×1 convolution for cross-channel interaction. The number of downsampling operations in the encoder is deliberately minimal (two blocks), matched by two upsampling passes in the decoder (bilinear interpolation). The output head is a 1×1 convolution mapping the final decoder features to three channels, followed by per-pixel softmax for class probabilities.

2. Optimization for Embedded, Real-Time Deployment

BioLite U-Net is explicitly optimized for low-latency inference and minimal resource use, targeting embedded platforms (such as the Raspberry Pi 4B) where GPU acceleration is typically absent. The architecture leverages depthwise separable convolutions, which reduce the number of multiply-accumulate operations and model parameters by a large factor.

Let $M$ be the number of input channels, $N$ output channels, $D_k$ kernel size, and $D_F$ feature map spatial dimension. The computational cost for a standard convolution is

$$C_{\text{standard}} = D_k \times D_k \times M \times N \times D_F \times D_F$$

while that for depthwise separable convolution is

$$C_{\text{depthwise}} = M \times D_k \times D_k \times D_F \times D_F + M \times N \times D_F \times D_F$$

which yields substantial savings for small $N$ and $D_k$.

The total parameter count is 0.01M (10,000 parameters), which is over 1300$\times$ smaller than a MobileNetV2-DeepLabV3+ baseline (13.35M parameters), and with FLOPs of only 0.44G versus 4.72G. Inference speed reaches 335 ms/frame on the Raspberry Pi 4B CPU without external accelerators.

3. Segmentation Performance and Quantitative Evaluation

BioLite U-Net is trained and evaluated using a manually annotated dataset of 787 RGB images (256 × 256 pixels) acquired during actual extrusion-based bioprinting sessions, with class annotations for nozzle, bioink, and background. The network is optimized for categorical cross-entropy loss on per-pixel softmax outputs.

Performance is benchmarked using mean Intersection over Union (mIoU), Dice score (spatial F1), and pixel-level accuracy. The reported figures are:

Model	mIoU (%)	Dice (%)	Accuracy (%)	Parameters (M)	Inference Time (ms)
BioLite U-Net	92.85	96.17	~99.55	0.01	335
MobileNetV2-DeepLabV3+	~94.33	97.03	~99.65	13.35	Not specified
MobileNetV3Small-FPN	91.68	95.52	~99.49	1.53	Not specified

BioLite U-Net achieves segmentation metrics that are competitive with, or closely approach, heavier MobileNet-based variants while utilizing orders-of-magnitude fewer resources. The model is robust in distinguishing nozzle, bioink, and background under varying illumination and process conditions—a task critical for closed-loop print quality assurance.

4. Dataset and Preprocessing

The underlying dataset consists of 787 RGB images, captured using a Raspberry Pi 1.6MP global shutter camera positioned beneath the bioprinting bed. Images are annotated manually with three semantic classes and split 80/10/10 into train/val/test sets. Preprocessing includes CLAHE for histogram normalization and geometric/photometric augmentation (rotation, flipping, brightness modulation) to promote generalization.

5. Advantages: Edge Deployment and Trade-off Analysis

BioLite U-Net is engineered for integration with real-world, on-device monitoring systems in bioprinting. Its model footprint (0.01M parameters) and compute profile permit deployment on microcontroller-class hardware without external acceleration, achieving sufficient throughput for bioprinting processes where real-time refers to update frequencies on the order of several frames per second.

While MobileNetV2-DeepLabV3+ achieves slightly higher scores in some metrics, the accompanying cost in memory and computation precludes its use on low-power embedded hardware. BioLite U-Net, by contrast, sustains practical segmentation accuracy (mIoU ≈ 93%, Dice ≈ 96%) at vastly reduced resource demand, confirming its suitability for closed-loop control scenarios.

6. Architectural Innovations and Broader Context

The use of depthwise separable convolutions is central to BioLite U-Net’s efficiency. Unlike standard U-Net implementations, which use regular convolutions in both encoder and decoder, this choice results in a model amenable to real-time embedded deployment while still retaining critical spatial detail via skip-connections—a design necessity given the importance of segmenting thin bioink filaments and nozzle tips.

The architecture is “shallow,” with only two downsampling blocks, in recognition of print monitoring’s relatively simple visual domain and the need to minimize latency. This design principle may be extended to related lightweight U-Net variants for other resource-constrained segmentation tasks.

7. Application in Intelligent Bioprinting Systems

Semantic segmentation delivered by BioLite U-Net supports intelligent, closed-loop bioprinting monitoring by classifying nozzle, bioink, and background during extrusion. With a throughput of 335 ms/frame on simple hardware, the architecture enables in situ feedback and fault detection (e.g., nozzle clogging or misalignment) at operational frequencies compatible with the requirements of bioprinting.

BioLite U-Net establishes a trade-off curve wherein moderate segmentation accuracy is delivered at exceedingly low resource cost, making it preferable for settings such as point-of-care diagnostics, portable lab instrumentation, and automation in tissue engineering environments.

Summary

BioLite U-Net represents a targeted lightweight segmentation architecture, constructed via depthwise separable convolutions and shallow encoder–decoder topology, optimized for in situ bioprinting monitoring on embedded edge hardware (Haider et al., 8 Sep 2025). This design achieves near state-of-the-art accuracy with a model footprint and computational profile that supports real-time, closed-loop process integration, marking a significant advance for deployable intelligent monitoring in biomedical manufacturing.