TernaryNet: faster deep model inference without GPUs for medical 3D segmentation using sparse and binary...

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Science

Background:

• Deep convolutional neural networks (DCNNs) are powerful tools in medical imaging, excelling at segmentation, localization, and prediction.
• However, their high computational demands limit their use in resource-constrained environments like mobile devices for point-of-care diagnostics.
• The need for efficient DCNNs that can operate without graphics processing units (GPUs) is critical for broader clinical adoption.

Purpose of the Study:

• To develop a novel scheme for approximating DCNN weights and activations using ternary values.
• To address the challenge of backpropagation with non-differentiable functions in DCNNs.
• To enable efficient DCNN inference without the need for GPUs, facilitating clinical applications.

Main Methods:

• Approximation of DCNN weights and activations using ternary values.
• Replacement of computationally expensive floating-point matrix multiplications with binary operators and population counts.
• Development of a ternary hyperbolic tangent continuation for backpropagation.

Main Results:

• Achieved over 90% memory reduction in a fully convolutional network for pancreas segmentation.
• Obtained high accuracy (71.0% Dice overlap) for pancreas segmentation, comparable to high-precision networks.
• Demonstrated sub-second inference times without GPUs, outperforming binary quantization methods.

Conclusions:

• The proposed ternary approximation technique is a key enabler for efficient DCNN inference without GPUs.
• This advancement can accelerate the integration of deep learning into practical clinical applications, including image-guided interventions and point-of-care diagnostics.
• The technique shows promise for enhancing accuracy in large-scale medical data retrieval.