A Framework for Endoscopic Image Classification Enhanced by Deep Transfer Learning

In this paper, we propose a framework for endoscopic image classification based on deep transfer learning (TL), specifically designed to address the unique challenges of endoscopic medical image classification. Our approach focuses on the performance of a convolutional neural network (CNN) model based on the ResNet architecture. To further improve the model's prediction accuracy and robustness, we introduce two ResNet variants: ResNe-SE and ResNet-CBAM, which incorporate the Squeeze-Excitation Module and Convolutional Block Attention Module, respectively. These modules allow the model to selectively learn significant features while suppressing noisy and unimportant features by capturing the dependencies between channels and spatial locations, ultimately achieving optimal performance compared to numerous baseline models. Furthermore, since limited training data can lead to overfitting of deep learning models, we apply deep TL to the field of medical image classification by fine-tuning parameters during model training based on models pre-trained on the publicly available ImageNet dataset. This approach addresses the problem of a limited number of endoscopic images. The results of our ablation experiments demonstrate the effectiveness of using deep TL techniques for this task, the improvements are 22.8%, 25.9%, and 13.3% for VGG-16, ResNet-34, and ResNet-50, respectively.