Lightweight skin cancer detection IP hardware implementation using cycle expansion and optimal computation arrays...

Area of Science:

• Medical Imaging
• Computer Engineering
• Artificial Intelligence

Background:

• Skin cancer is a global health threat requiring accurate early diagnosis.
• Medical image classification for skin lesions faces challenges in deployment on low-power devices.
• Computational demands of current algorithms hinder efficiency and energy conservation.

Purpose of the Study:

• To propose a lightweight hardware design for real-time skin disease classification on portable devices.
• To develop an efficient and low-power solution for skin cancer detection using convolutional neural networks (CNNs).
• To address the limitations of existing algorithms in terms of computational complexity and resource utilization.

Main Methods:

• A hardware design based on CNNs with an optimized parallel processing engine (PE).
• Implementation of loop unrolling to reduce data accesses and computational complexity.
• Utilizing 16-bit floating-point numbers for data inference in convolutional, pooling, and fully connected layers.
• All-hardware FPGA-based implementation for skin cancer detection.

Main Results:

• Achieved an average classification accuracy of 97.8% on the HAM10000 dataset.
• Demonstrated a 3.5x speedup in recognition compared to existing accelerators at 50 MHz.
• Consumed only 0.48 W of power, meeting portable device constraints.
• Showcased low resource utilization and cost-effectiveness.

Conclusions:

• The proposed hardware architecture effectively meets the constraints of portable devices for skin cancer detection.
• This all-hardware FPGA-based platform offers efficient classification, high accuracy, low power consumption, and cost-effectiveness.
• The design represents a significant advancement in real-time, on-device medical image analysis for dermatology.