With the progressive growth of artificial intelligence (AI), the recent advancements in deep learning and applying the latest deep learning methods for medical imaging have become an active research area both in the medical industry and academia. Newly advanced deep learning technologies have successfully solved various medical imaging problems with high accuracy, efficiency, stability, and scalability.
Owing to the significant breakthroughs attained by deep learning models, it is envisioned that the interaction between patients and AI-based medical systems will enable soon with high safety and convenience, such intelligent systems will genuinely ameliorate patient healthcare. The core traits of medical imaging, namely images-multimodality & dense, tasks-diverse & complex, samples-imbalanced & heterogeneous, labels-sparsity & noisy, data-non standard & isolated, and diseases-numerous & long-tailed which are contend using AI-based technologies.
Key Technologies in Medical imaging and Deep Learning:
Various key technologies arise from the diverse medical imaging with deep learning applications, including medical image reconstruction - medical image enhancement - medical image segmentation - medical image registration - medical image classification - Computer-Aided Detection(CAD) - Other technologies include landmark detection, image or view recognition, automatic report generation.
Modern Deep Learning Approaches for Medical Imaging:
As deep learning resembles a multilayered human cognition system, it has gained remarkable attention in medical imaging applications. Several deep learning approaches emerge to exploit various kinds of medical imaging analysis and applications. Some of the supervised deep neural architectures utilized for medical imaging are Convolutional Neural Network (CNN), Pre-trained CNN, and Recurrent Neural Network (RNN). Few other unsupervised deep learning architectures are Autoencoder (AE), Deep Belief Network (DBN), and Generative Adversarial Network (GAN).
Adversarial and attention mechanisms - Adversarial learning is broadly used in medical imaging, for instance, medical image reconstruction, image quality enhancement, and segmentation. The attention mechanism permits automatic discovery of reason and location to focus on describing medical image contents and making an exhaustive decision. Thus, the attention mechanism is integrated with GAN for medical image analysis.
Neural Architecture Search (NAS) - Deep learning proficiently applies NAS to volumetric medical image segmentation. The design of the architecture of the NAS deep network provides high performance for medical image applications.
Transfer learning (TL) - TL utilizes the trained deep network and fine-tunes it for medical imaging tasks, conducive to speeding up training convergence and boosting accuracy. Currently, various medical image challenges with diverse modalities, target organs, and pathologies learn 3D networks with an effective pretrained model for 3D medical image analysis tasks.
Domain adaptation - The domain adaptation module maps the target input to features based on the alignment of the source domain feature space and domain critic module to discriminate the feature space of both domains, developed for cross-modality biomedical image segmentation. Presently, an integrated learning mechanism with a Universal network (U Net) proffers a new potentiality in dealing with multiple domains and even multiple heterogeneous tasks in medical imaging.
Self-supervised learning - Self-supervised framework uses a representative task for the recovery of the original image from the deformed image. Possible distortions in the medical image include non-linear gray-value transformation, local pixel shuffling, and image out-painting and in-painting. This learning framework recently permits the network via cube proxy task for learning invariant to translation and rotation features and sturdy to noise.
Semi-supervised learning - Semi-supervised learning is developed for cardiac image segmentation by training a deep learning model using a less set of annotated images. Currently, the attention-based semi-supervised deep network for segmentation is designated that adversarially trains a segmentation network and computes a confidence map as a region attention-based semi-supervised learning strategy to cover the unlabeled data.
Weakly or partially supervised learning - Weakly-supervised learning paradigm is highly utilized for multi-label disease classification based on image-level or weak annotations. For multi-organ segmentation, this learning strategy learns a single multi-class network from multiple datasets with partial organ labels. This strategy was also applied to detect abnormal regions in the diseased images using the deep learning model.
Unsupervised learning - In medical imaging, unsupervised learning and disentanglement have been widely exploited for image registration, motion tracking, artifact reduction, improving classification, domain adaptation, and general modeling by promoting the statistical matching of deep features without the existence of annotated images.
Federated learning - Federated learning for medical imaging mainly focus on data privacy, data security, data access rights, robust algorithmic model via distributed computing, and model aggregation strategies without transferring data outside a hospital or an imaging lab. Recently, federated learning has been utilized for multi-institutional deep learning model without patient data and report sharing for brain lesion segmentation, data privacy for brain tumor segmentation, and discovery of disease-related biomarkers.
State-of-the-art Applications of Deep Learning in Medical Imaging:
Several clinical practices in medical imaging promote deep learning, including chest, neurology, mammography, cardiovascular, gastroenterology, dermatology, abdominal, pulmonary, and microscopy imaging applications. Some of the leading-edge medical imaging applications are discussed below;
Thoracic imaging: Deep learning in thoracic imaging conducts segmentation of anatomical structures, detection, and diagnosis in chest radiography, decision support in lung cancer screening and COVID-19 case study and monitoring systems with the assistance of Computed Tomography (CT) scans.
Neuroimaging: Latterly, deep learning attains a drastic rise in famousness within the neuroimaging community. Neuroimage segmentation and tissue classification, deformable image registration, and neuroimaging prediction are the neuroimaging applicative tasks of deep learning. Deep learning in neuroimaging currently exploits GAN, optimal hyperparameter selection, domain adaptation, and semi-supervised designs.
Cardiovascular imaging: With the recent progress of data-driven deep learning, the quantification and understanding of cardiac anatomy and function via cardiovascular imaging brings great impact. Deep learning models such as graph convolutional network, CNN, fully convolutional network, and RNN are employed for cardiovascular imaging tasks, including cardiac image segmentation, cardiac motion tracking, cardiac vessel segmentation, plaque characterization, and coronary artery segmentation.
Abdominal imaging: Deep learning is increasingly applied for automated detection, classification, and segmentation that relies on abdominal anatomy and disease using medical imaging. Localization and segmentation of organs and lesions and opportunistic screening are the recent interest in abdominal image analysis via deep learning.
Microscopy imaging: In digital pathology, deep learning approaches are applied to predict the basic molecular and mutational status of the disease and cells. More recently, deep learning algorithms have been interesting in microscopic level pathology applications such as nuclei detection and segmentation, disease grading, quantitative analysis of cells in live-cell imaging experiments, mutation identification and its pathway association, survival, and disease outcome prediction using microscopy images.
Challenges and Future Perspectives of Deep Learning-based Medical Imaging:
The idea of applying deep learning-based algorithms for medical imaging is an engrossing and booming research area; however, several complexities reduce its progress. Some of the big barriers such as unavailability of the annotated dataset, privacy and legal issues, dedicated medical experts, and nonstandard data are demanding high research works on deep learning for effective medical imaging solutions. A few of the future promises that need to focus on are discussed below:
Extensive inter-organization collaboration - Collaboration between hospital providers and deep learning research scientists is required to solve data unavailability, which is beneficial to providing more sophisticated deep learning techniques for healthcare data.
Capitalize on big image data - annotation of medical image data is expensive and time-consuming. Thus in the future, sharing the data resource among various healthcare service providers will be the better solution.
Advancement in deep learning methods - There is a need to shift from supervised deep learning in the medical image analysis field to unsupervised or semi-supervised learning concepts that possess unlimited opportunities for improvement.
Privacy and legal issues - The prevailing deep learning methods need to be developed, as anonymizing the patient information to prevent its disclosure is a big challenge.