Supervised Contrastive Learning Framework and Hardware Implementation of Learned ResNet for Real-Time Respiratory...

Area of Science:

• Medical device technology
• Machine learning for healthcare
• Signal processing for diagnostics

Background:

• Respiratory sound classification is crucial for diagnosing pediatric conditions.
• Data scarcity and class imbalance pose significant challenges in training diagnostic models.
• Real-time monitoring requires efficient and low-latency computational hardware.

Purpose of the Study:

• To develop a supervised contrastive learning (SCL) framework for robust respiratory sound classification.
• To implement a learned ResNet model on a field-programmable gate array (FPGA) for real-time monitoring.
• To optimize algorithmic and hardware aspects for efficient and accurate respiratory sound analysis.

Main Methods:

• Utilized supervised contrastive learning (SCL) with feature augmentation and MixUp to address data limitations.
• Employed Bayesian optimization for hyperparameter tuning in pre-processing and SCL.
• Implemented algorithm-hardware co-optimizations including Quantization-Aware Training (QAT) and network layer merging for FPGA deployment.

Main Results:

• Achieved a total score of 0.8725 on the ResNet-18 model for multi-class classification tasks using the SPRSound dataset.
• Reduced model size by 40% and computation latency by 70% through hardware optimizations.
• Deployed the ResNet model on a Xilinx Zynq ZCU102 FPGA with 16ms latency and minimal inference degradation (<2%).

Conclusions:

• The proposed SCL framework effectively handles data scarcity and class imbalance in respiratory sound classification.
• Algorithm-hardware co-optimization enables efficient real-time respiratory monitoring on FPGA with high accuracy.
• This integrated approach demonstrates a viable solution for portable and accurate pediatric respiratory diagnostics.