Cutting-edge artificial intelligence techniques especially deep learning algorithms have shown great potentials in data-driven diagnostics. Convolutional neural networks (CNNs) have been widely applied in image analysis, pattern recognition, and anomaly detection. CNNs can automatically learn features from images, avoiding human bias and improving the efficiency. The multi-layer deep network structure enables CNN to extract features at different abstraction levels in images, enhancing semantic information in images and improving its performance in various tasks such as classification, segmentation, and detection. CNN exhibits great potentials in the diagnosis, prognosis and classification of various diseases. Whereas, there are some unmet challenges in data quality and quantity, data security and privacy, model interpretability, and ethical considerations. This chapter summarizes the advantages and challenges of the state of the art, and future directions under the context of healthcare 5.0, providing a reference for clinical researchers, data scientists, and biomedical engineers.
TopIntroduction
Cutting-edge artificial intelligence (AI) techniques are reshaping the ecosystem of diagnostics. Deep learning algorithms consist a major branch of AI, including convolutional neural networks (CNNs), recurrent neural networks (RNNs), autoencoders, generative adversarial networks (GANs), and reinforcement learning (Alzubaidi et al., 2021). These methods excel in handling complex medical data, enhancing diagnostic accuracy, and reducing costs and risks. CNNs shine as stars in image analysis, pattern recognition, and anomaly detection. They have shown outstanding performance in healthcare applications such as cancer, ophthalmic diseases, and lung diseases.
As illustrated in Figure 1, a CNNs consists of sever key components, including an input layer, a convolution layer, a pooling layer, an activation function, a fully connected layer, and an output layer (Bhatt et al., 2021). This multilayer deep network structure of CNN provides unique advantages in healthcare applications. First, CNNs can automatically learn features from images without the need for manual design or feature extractors. This eliminates human bias and limitations, improving the quality and diversity of features. Second, convolutional layers can be used to extract local features from images, effectively compressing and reducing the dimensionality of imaging data. This reduces computational complexity and memory consumption, improving operational efficiency. Third, pooling layers can be used to achieve down sampling of features in the image by merging semantically similar features into one, thus providing invariance to translation, rotation, scaling, and other changes in the image. This enhances the network's robustness and generalization ability, improving pattern and anomaly recognition. Finally, stacking multiple convolutional layers can construct deep network structures, allowing the extraction of features at different levels and abstraction levels within the image. This enhances semantic information in the image and improves the performance of tasks such as classification, segmentation, and detection.
Recent years have witness rapid growth in the application of CNNs in different clinical settings. To understand the state of the art and future directions, we systematically reviewed recent research works and analyzed the challenges and future directions under the context of healthcare 5.0. This chapter provides a reference for clinical researchers, data scientist, and biomedical engineers.
Figure 1. The components of a convolutional neural network (CNN) (adapted from Bhatt et al., 2021 under a Creative Commons Attribution 4.0 International License [CC BY 4.0])
Literature search strategy
In literature research, a set of keywords were used, including “artificial intelligence,” “disease detection”, “computer vision”, “deep learning”, “convolutional neural networks”, and “medical data analysis”. These keywords and their synonyms were combined to conduct searches in mainstream databases including Web of Science, PubMed, Google Scholar, and Scopus. All the items published in English in recent five years (from 2018 to 2023) were selected for screening.
We carefully reviewed the abstracts of all the selected papers and excluded those irrelevant to CNNs. The finally selected papers were categorized according to the sub-disciplines within the healthcare fields. We conducted a comparative analysis that encompassed various aspects, including the subjects of study, study design, algorithm optimizations, data sources, prediction accuracies, and contributions. We also briefly reviewed the history of CNN. The goal was to identify recent trends in the application of CNNs in healthcare, important considerations during implementation, as well as the roles and potentials of CNN models in modern healthcare technologies towards healthcare 5.0.