DLA-H: A Deep Learning Accelerator for Histopathologic Image Classification

Area of Science:

• Computer Science
• Medical Imaging
• Hardware Acceleration

Background:

• Deep learning, particularly deep neural networks, is widely applied in image classification, including medical sciences.
• High computational demands of deep learning pose challenges in power consumption and runtime.
• Deep learning accelerators are crucial for enhancing performance and energy efficiency.

Purpose of the Study:

• To propose a novel deep learning accelerator (DLA-H) and its data flow (BJS) optimized for histopathologic image classification.
• To address the computational challenges and improve energy efficiency in deep learning for medical image analysis.

Main Methods:

• Design of a specialized deep learning accelerator architecture (DLA-H).
• Development of an efficient data flow (BJS) emphasizing data reuse and localization.
• Simulation and performance evaluation using the MAESTRO tool.

Main Results:

• The proposed DLA-H and BJS achieved a total runtime of 756 cycles.
• Demonstrated [Formula: see text] GFLOPS roofline throughput, indicating extreme performance improvement.
• Significant gains in performance and energy efficiency compared to general-purpose accelerators.

Conclusions:

• The DLA-H accelerator and BJS data flow offer a substantial advancement for histopathologic image classification.
• The proposed solution effectively tackles the computational bottlenecks in deep learning for medical imaging.
• This work contributes to more efficient and powerful deep learning hardware for scientific applications.