An experimental study of deep CNN models for medical image classification

Abstract

Deep learning, a subset of machine learning, has revolutionized artificial intel- ligence by enabling models to learn complex patterns directly from data. Its application in medical image analysis aids in early detection, diagnosis, and ef- fective treatment planning, significantly improving patient outcomes. Accurate classification of medical images is crucial for timely medical interventions. This project leverages deep learning techniques to classify medical images from three datasets: lung cancer, breast cancer, and kidney cyst/tumor conditions. We em- ploy five state-of-the-art Convolutional Neural Network (CNN) models—VGG16, Xception, EfficientNetB0, NasNetMobile, and DenseNet121—on each dataset. To ensure optimal performance, each model undergoes rigorous hyperparameter tun- ing and is evaluated using metrics like F1 score, accuracy, precision, and recall. After establishing baseline performances, we refine the models by incorporating the Convolutional Block Attention Module (CBAM) into the last convolutional layer of each CNN. CBAM enhances the model’s focus on relevant image parts by adaptively recalibrating feature responses, addressing challenges such as subtle differences between tissues, class imbalance, and noise. Our results show sig- nificant improvement in classification accuracy and overall performance metrics with CBAM integration. The enhanced models consistently outperform their non- attention counterparts across all datasets. The findings highlight the potential of attention mechanisms like CBAM in improving the accuracy and reliability of deep learning models for medical image analysis. This integration boosts performance and provides a more interpretable framework for understanding neural network decisions, contributing to the development of reliable diagnostic tools for better patient outcomes. By addressing inherent challenges in medical image classifica- tion and demonstrating the effectiveness of advanced deep learning techniques, this project underscores the transformative impact of deep learning in healthcare, leading to more precise and efficient diagnostic processes.