Technical Summary: BioLite U-Net for Edge-Deployable Bioprinting Monitoring

Problem Statement

Bioprinting offers a transformative approach to fabricating tissue and organ models but faces critical challenges in ensuring the fidelity and consistency of printed structures in real-time. Effective in situ monitoring is hindered by two primary constraints:

1. Dynamic Complexity: The extrusion process involves rapid nozzle movement, variable flow rates, and deformable bioinks, requiring high spatial and temporal resolution to capture relevant features.
2. Resource Constraints: Most bioprinters operate in environments with limited computational resources, making the deployment of conventional deep learning models (which rely on high-end GPUs) impractical.

While semantic segmentation is widely used in fields like autonomous driving and medical imaging, its application to bioprinting on resource-constrained embedded devices remains underexplored. Existing methods often rely on simple classification or thresholding heuristics that lack dense semantic labeling, failing to provide the granular feedback necessary for closed-loop control.

Methodology

Dataset Creation

The authors introduced a novel, manually annotated dataset comprising 787 RGB images captured during real-time bioprinting operations.

• Capture Setup: Images were taken using a Raspberry Pi 1.6-megapixel global shutter camera with a macro lens, positioned below the printing bed to focus on the extrusion needle tip.
• Classes: The dataset is labeled into three semantic classes: Background (printer frame, bed, environment), Bioink (extruded filament), and Nozzle (dispensing tip).
• Variability: Data was collected under ambient lighting with varied extrusion speeds, nozzle geometries, and alginate-based bioink types to ensure robustness.
• Preprocessing: Images were resized to 256×256. Training utilized augmentations including horizontal flipping, random brightness/contrast adjustments, and rotation (±15°).

Model Architecture: BioLite U-Net

The core contribution is a lightweight semantic segmentation framework named BioLite U-Net, designed specifically for edge deployment.

• Architecture: An encoder-decoder topology with skip connections, inspired by the original U-Net but optimized for efficiency.
• Key Optimizations:
  • Depthwise Separable Convolutions: All standard convolutions are replaced with depthwise separable convolutions (3×3 depthwise followed by 1×1 pointwise) to drastically reduce parameters and computational load.
  • Shallow Encoding/Decoding: The network utilizes few downsampling and upsampling layers (using bilinear interpolation) to balance feature richness with low latency.
  • Skip Connections: Feature maps from the encoder are concatenated with decoder maps at each resolution to preserve spatial details essential for thin structures like ink traces.
• Output: A final 1×1 convolution maps features to three channels, followed by a softmax layer for per-pixel class probability.

Training and Evaluation

• Training: Models were trained in PyTorch using the Adam optimizer (learning rate $10^{-3}$, weight decay $10^{-5}$) for 200 epochs with a batch size of 4. Categorical cross-entropy loss was used.
• Baselines: BioLite U-Net was compared against MobileNetV2-DeepLabV3+ and MobileNetV3-small-FPN.
• Deployment: All models were evaluated on a Raspberry Pi 4B (1.5 GHz quad-core CPU, 4 GB RAM) without GPU acceleration or quantization to assess real-world feasibility.

Key Contributions

1. Novel Dataset: The first publicly available, manually annotated dataset for bioprinting semantic segmentation, addressing a gap in the literature regarding semantic understanding of the extrusion process.
2. BioLite U-Net Architecture: A custom, lightweight architecture leveraging depthwise separable convolutions to achieve real-time inference on low-power devices without compromising segmentation accuracy.
3. Edge Deployment Benchmark: A comprehensive evaluation demonstrating the trade-off between performance and efficiency on embedded hardware, establishing a baseline for intelligent, closed-loop bioprinting systems.

Results

Quantitative Performance

On the test set, BioLite U-Net achieved:

• mIoU: 92.85%
• Dice Score: 96.17%
• Pixel Accuracy: 99.55%

While MobileNetV2-DeepLabV3+ achieved slightly higher mIoU (94.33%) and Dice (97.03%), BioLite U-Net offers a superior trade-off regarding model size and computational cost.

Efficiency and Complexity

• Parameters: BioLite U-Net contains only 0.01 million parameters, making it over 1,300× smaller than MobileNetV2-DeepLabV3+ (13.35M) and 150× smaller than MobileNetV3Small-FPN (1.53M).
• FLOPs: The model requires 0.44 GFLOPs per inference, approximately 10–11× fewer than the baselines.
• Inference Latency: On the Raspberry Pi 4B, BioLite U-Net processes a frame in 335 ms. While slower than the baselines on CPU (MobileNetV3Small-FPN took 260.9 ms), the authors note that lower FLOPs do not always translate to lower CPU latency due to memory copy and cache miss penalties on edge devices. However, the 335 ms latency is deemed sufficient for bioprinting feedback, where extrusion dynamics occur over seconds rather than milliseconds.

Qualitative Results

The model successfully segments fine structures and preserves boundaries under varying illumination. Minor mis-segmentation was observed in low-contrast regions, but the model generally matched ground truth labels effectively, including in challenging areas like filament transitions.

Significance and Claims

The paper claims to present the first study to perform real-time semantic segmentation of bioprinting imagery on embedded hardware. The significance of this work lies in:

• Enabling Closed-Loop Control: By providing accurate, on-device segmentation, the system can trigger corrective actions for nozzle misalignment or extrusion faults, enhancing print quality and biological viability.
• Resource Efficiency: The model fits within the memory and compute constraints of commodity embedded boards (like the Raspberry Pi 4B) without requiring external accelerators or quantization.
• Practical Feasibility: The inference time of ~335 ms is validated as suitable for bioprinting monitoring, aligning with the temporal scale of the physical process.

The authors conclude that BioLite U-Net establishes a foundation for smart, autonomous bioprinters capable of adaptive control, moving the field toward scalable and precise tissue fabrication. Future work is identified as extending to multi-material printing, sensor fusion (e.g., RGB-IR), and temporal modeling to further enhance robustness.