TernScope: Model-Hardware Co-Design for Efficient Ternary Endoscopic Segmentation
Summary:TernScope introduces a novel model-hardware co-design approach for efficient endoscopic segmentation in edge computing environments. By learning a compact segmentation network with ternary weights and 8-bit activations, TernScope reduces computational complexity and hardware execution overhead. The model employs split-channel large-kernel blocks and branch fusion to preserve spatial context while minimizing redundant inference paths. Ternary Progressive Quantization-aware Training (TPQ) further
Technical Mechanism Analysis
TernScope achieves a high-efficiency model-hardware co-design through the following key techniques:
-
Ternary Weights and 8-bit Activations: All 25 convolutional layers use ternary weights (-1, 0, 1) and 8-bit activations. This not only reduces the model parameter count but also simplifies hardware multiplication operations, enabling multiplier-free computation.
-
Split-Channel Large-Kernel Blocks: The split-channel design allows large-kernel blocks to preserve spatial context while reducing computational complexity. This is particularly beneficial for medical image segmentation tasks where spatial information is crucial.
-
Branch Fusion: By fusing redundant inference paths, TernScope further reduces computational overhead and improves inference efficiency.
-
Ternary Progressive Quantization-aware Training (TPQ): The TPQ method gradually adjusts the network from deep to shallow layers to adapt to quantized weights and activations. This ensures that the impact of quantization on model accuracy is minimized.
Engineering Trade-offs and Performance
TernScope achieves a good balance between model accuracy and hardware efficiency. Despite having only 0.2483 million parameters and 0.0949 GMACs, it maintains competitive segmentation accuracy on both in-domain and external datasets. Compared to the traditional Gemmini (a systolic-array accelerator with 8-bit multipliers), TernScope reduces the compute-array area by 88.46% under the same 45nm technology.
However, this design also involves some trade-offs. For example, the use of ternary weights, while reducing storage and computation requirements, may limit the model's expressive power. Additionally, the TPQ training process increases training time, but this is a necessary trade-off to ensure the performance of the quantized model.
Developer Deployment Recommendations
For developers, TernScope's hardware-friendly design makes it highly suitable for deployment on resource-constrained edge devices. Here are some recommendations:
-
Hardware Optimization: It is recommended to implement TernScope on programmable hardware such as FPGAs or ASICs to take full advantage of its low-bit operation benefits.
-
Quantization-aware Training: Before deployment, use the TPQ method to perform quantization-aware training on the model to ensure its performance after quantization.
-
Model Compression: Combining other model compression techniques (such as pruning and knowledge distillation) can further reduce model size and improve inference efficiency.
Conclusion
TernScope demonstrates the great potential of model-hardware co-design in edge medical segmentation. Its efficient design and strong performance make it an ideal choice for medical applications on low-power devices.
— END —Source: Hugging Face Daily Papers (2026-10-21)
Tags: #TernScope #Model-Hardware Co-Design #Medical Segmentation #Edge Computing #Quantization-aware Training
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments