TernScope: Model-Hardware Co-Design for Efficient Ternary Endoscopic Segmentation

TernScope. Model-Hardware Co-Design for Efficient Ternary Endoscopic Segmentation. By Zhuoyu Wu and five others, from Monash University Malaysia, the Hong Kong University of Science and Technology, and two other institutions. Published on the twenty-first of October.

Medical segmentation at the edge has to reconcile three demands that pull against one another: a small model, high accuracy, and fast inference. Shrink the model and accuracy is at risk. Protect the accuracy and the model may stay too large or too slow. A design has to give ground somewhere, or find a way to make the demands agree.

TernScope treats that tradeoff as a co-design problem spanning the model and the hardware. It learns a compact segmentation network whose low-bit operators also simplify how the hardware executes it, so a saving on one side becomes a saving on the other.

The network is built from split-channel large-kernel blocks, which preserve spatial context with limited computation, while branch fusion removes redundant inference paths. Ternary Progressive Quantization-aware Training, or TPQ, then adapts the fused network in a set order: from deep layers to shallow ones, through the decoder, and finally to the input and prediction layers. By the end, all twenty-five convolutions use ternary weights and eight-bit activations. That enables packed storage and computation with no multipliers.

The resulting model has zero point two four eight three million parameters and needs zero point zero nine four nine GMACs, while keeping competitive segmentation accuracy on in-domain and external datasets. On the hardware side, under the same forty-five nanometre technology, TernScope reduces compute-array area by eighty-eight point four six percent compared with the original Gemmini, a systolic-array accelerator built with eight-bit by eight-bit multipliers. Both halves of the co-design are measured, then: the model by its size, compute and accuracy, the hardware by the area it occupies.

The authors present this as a path from accurate lightweight medical segmentation to compact low-bit hardware. The two halves meet at the ternary weights. The same choice that shrinks the model is the one that lets the accelerator do without its multipliers.