Grad-CAM Visualization and Ensemble Learning for Improved Gastrointestinal Disease Classification Using CNNs
Chih Mao Tsai, Jiann-Der Lee · 2024
The aim of this paper is to investigate the efficacy of ensemble learning techniques in convolutional neural network (CNN) models for gastrointestinal disease classification using the Kvasir v2 dataset [1]. A particular emphasis is placed on visualizing and comprehending model predictions through Gradient-weighted Class Activation Mapping (Grad-CAM). Eleven pretrained CNN models such as VGG16, ResNet50, NASNetMobile, NASNetLarge, MobileNetV3Large, InceptionV3, InceptionResNetV2, EfficientNetB4, DenseNet201, DenseNet121, and VGG19 were fine-tuned on the Kvasir v2 dataset. These models were assessed based on accuracy, precision, recall, F1 score, and specificity. The top-performing models were further combined using ensemble methods, including hard voting and weighted averaging, to improve classification performance. Grad-CAM was employed to produce visual explanations of model predictions, enabling a detailed analysis of the impact of different ensemble combinations on the decision-making process. This study demonstrates that ensemble learning can enhance model performance and provides valuable insights into the interpretability of CNN models in medical image analysis.