Early Exit Based on Deep Learning Model for Polyp Colonoscopy Image Classification

Early exit is a widely adopted approach to reduce the inference time of deep learning models. By introducing side-branch classifiers into the main backbone network, this approach allows test samples to be predicted and exit the network early when high confidence is achieved. While the early exit mechanism has been extensively explored in various computer vision applications, its use in medical imaging remains relatively underexplored. In this study, we propose to design a lightweight early exit branch for polyp colonoscopy image classification with a combination of Convolutional Block Attention Module (CBAM) and Fully Connected Layer (FC). These branches are embedded into a deep learning backbone to leverage intermediate features for early predictions. Extensive experiments on the Kvasir polyp dataset demonstrate that our method achieves a favorable trade-off between accuracy and computational efficiency, showcasing its potential of lightweight early exit mechanisms to improve the efficiency of deep learning systems in medical image analysis, paving the way for faster and more resource-efficient diagnostic tools.