A review of deep learning in medical imaging: Imaging traits, technology trends, case studies with progress highlights, and future promises (2008.09104v2)

Abstract: Since its renaissance, deep learning has been widely used in various medical imaging tasks and has achieved remarkable success in many medical imaging applications, thereby propelling us into the so-called AI era. It is known that the success of AI is mostly attributed to the availability of big data with annotations for a single task and the advances in high performance computing. However, medical imaging presents unique challenges that confront deep learning approaches. In this survey paper, we first present traits of medical imaging, highlight both clinical needs and technical challenges in medical imaging, and describe how emerging trends in deep learning are addressing these issues. We cover the topics of network architecture, sparse and noisy labels, federating learning, interpretability, uncertainty quantification, etc. Then, we present several case studies that are commonly found in clinical practice, including digital pathology and chest, brain, cardiovascular, and abdominal imaging. Rather than presenting an exhaustive literature survey, we instead describe some prominent research highlights related to these case study applications. We conclude with a discussion and presentation of promising future directions.

A Review of Deep Learning in Medical Imaging: Imaging Traits, Technology Trends, Case Studies with Progress Highlights, and Future Promises

Introduction

The surveyed paper provides a comprehensive overview of the utilization of deep learning (DL) in medical imaging, highlighting its significant impact across various modalities and tasks. This review addresses the intrinsic traits of medical imaging, technological advancements in DL architectures, and practical applications through selected case studies. The authors meticulously outline challenges such as sparse labeling and distribution drift and propose emerging solutions like federated learning and annotation-efficient approaches. They further explore the intersection of medical AI with clinical practice and speculate on future advancements.

Key Traits of Medical Imaging

Medical imaging is characterized by its high data density and variability across modalities such as CT, MRI, and ultrasound. The heterogeneity in disease patterns and imaging protocols poses a substantial challenge to DL approaches. Sparse and noisy labels further complicate model training processes. The review emphasizes these challenges in the context of DL applications, advocating for innovative methodologies to address these constraints.

Emerging Trends in Deep Learning

The authors identify several key technological trends that have emerged to tackle the aforementioned challenges:

• Network Architectures: Advances in architectures such as U-Net, ResNet, and DenseNet have revolutionized image segmentation and classification, paving the way for robust medical imaging solutions.
• Annotation-Efficient Approaches: Strategies like transfer learning, domain adaptation, and semi-supervised learning are crucial in addressing label scarcity while maintaining model accuracy and generalization.
• Federated Learning: This approach mitigates privacy concerns by enabling decentralized model training across institutions, maintaining data integrity while improving model robustness.
• Interpretability and Uncertainty Quantification: These elements are vital for gaining clinical trust and ensuring that DL models provide actionable insights, aligning with human interpretability standards.

Case Studies

The paper presents focused case studies that demonstrate the utility and progress of DL in specific medical imaging domains:

• Thoracic Imaging: DL has made substantial strides in the segmentation of anatomical structures and the detection of pathologies like COVID-19 and lung cancer, supported by techniques like CO-RADS for CT severity scoring.
• Neuroimaging: Applications in neuroimaging include brain age prediction and Alzheimer's disease modeling, utilizing deep networks for segmentation and registration tasks effectively.
• Cardiovascular Imaging: The integration of motion tracking and segmentation technologies has aided in the detailed analysis of cardiac function, leveraging novel architectures for improved outcomes.
• Abdominal Imaging: Automated organ segmentation, lesion detection, and opportunistic screening represent areas where DL has significantly enhanced diagnostic abilities.
• Microscopy Imaging: DL aids in the detection and classification of nuclei and the prognosis of diseases by predicting genetic mutations from tissue morphology.

Discussion

The review outlines persistent challenges, such as developing systems robust to cross-domain variability and achieving clinical acceptance. The authors call for improved integration of DL tools into clinical workflows, advocating for frameworks that demonstrate proven clinical benefits. They foresee the potential of DL to extend into holistic patient-level analysis, integrating multi-modal data for comprehensive patient assessments and personalized treatment plans.

Conclusion

This paper serves as a crucial resource for researchers and practitioners, detailing the current state and future potentials of DL in medical imaging. By addressing the complex interplay of technical challenges and clinical applications, the authors illuminate pathways for continued innovation in AI-driven healthcare solutions.