Ensemble Architectures and Efficient Fusion Techniques for Convolutional Neural Networks: An Analysis on Resource Optimization Strategies

• Cícero L. Costa IFMT / UFU https://orcid.org/0000-0003-3789-5074
• Danielli A. Lima IFMT https://orcid.org/0000-0003-0324-6690
• Celia A. Zorzo Barcelos UFU
• Bruno A. N. Travençolo UFU https://orcid.org/0000-0001-7690-301X

Resumo

The human gastrointestinal tract is prone to various abnormalities, including lethal diseases such as cancer, necessitating better endoscopic performance and standardized screening. Endoscopic scoring systems lack generalizability, emphasizing the need for artificial intelligence-based solutions. Using the HyperKvasir dataset, we employed deep learning, specifically Convolutional Neural Networks, or shortly CNNs, to analyze endoscopic images and videos. Our study focused on improving the classification of gastrointestinal tract diseases by proposing various CNN ensembles and fusion techniques. Through the use of seven CNN models and effective merging techniques, we achieved enhanced performance. Validation involved literature review and experiments. DenseNet-161 influenced the merger process, and integrating ResNet152 and VGG further enhanced effectiveness. Resource analysis included GPU model, RAM usage, and execution time. Results demonstrated comparable performance to the previous model, with F1-score of 0.910 and Matthews correlation coefficient, MCC for short, of 0.902, using 10 GB GPU RAM (compared to 15.8 GB). With 24.7 GB GPU RAM, F1-score of 0.913 and MCC of 0.905 were achieved. These findings advance our understanding of ensemble architectures and fusion techniques.