Effect of Post-Training Pruning and Quantization on Endoscopic Computer-Aided Diagnosis Models

Computer-aided diagnosis (CADx) in the endoscopy domain has generated significant interest, driven by advances in computer hardware over the past two decades. However, current state-of-the-art Deep Neural Network (DNN) architectures are overparameterized and often exhibit a multitude of redundancies, leading to computational inefficiencies, hindering real-time video analysis and deployment on resource-constrained devices. In this research, we evaluate the efficacy of post-training DNN compression techniques, pruning and quantization, within the realm of CADx systems for endoscopic imaging. The results demonstrate, on average, pruning and quantization can reduce memory size to only 11.06% of the employed base model and achieve a 4.22 times faster inference speed, which is essential for real-time clinical implementation. Our findings highlight the potential of network compression to optimize deep learning models for endoscopy, enabling more efficient and practical utilization in the clinic.