Pre-Trained Lightweight Deep Learning Models for Surgical Instrument Detection: Performance Evaluation for Edge Inference

Surgical instrument detection and tracking is essential in surgical procedures. The presence and proper use of surgical instruments is crucial for patient safety and overall patient outcomes. Most existing deep learning-based surgical instrument detection systems require high-end GPUs or cloud services for deployment. However, such systems are not feasible in resource-constrained or remote areas where access to expensive and bulky equipment is limited. To overcome these challenges, we propose a cost-effective and sustainable solution for deep learning based surgical instrument detection. Our approach is to detect surgical instruments from images using pretrained models on a Raspberry Pi. We compare two lightweight, pretrained deep learning models, EfficientNet and MobileNet, on their accuracy and efficiency for this task. These models are optimized for edge devices with limited computational resources. We evaluate the tradeoff between model accuracy and computational efficiency on a benchmark dataset. We discuss the strengths and limitations of each model, as well as their implications for developing surgical instrument detection applications for edge devices. This work can enable the selection of effective deep learning models for real-time inference on edge devices, and facilitate the development of efficient and cost-effective healthcare solutions.