한국어

Chapter 6: AI and Machine Learning

Artificial Intelligence in Medical Imaging Diagnostics

6.1 AI Revolution in Medical Imaging

Artificial intelligence is transforming medical imaging from traditional pattern recognition to automated analysis, detection, and diagnosis. Deep learning algorithms can now match or exceed human performance in specific imaging tasks.

700+
FDA-Cleared AI Devices
$2.5B
Medical Imaging AI Market
94%
Top Algorithm Accuracy
30%
Potential Time Savings

6.2 Deep Learning Fundamentals

Deep learning neural networks form the foundation of modern medical imaging AI:

Key Architectures

Architecture Description Application
CNN (Convolutional) Hierarchical feature extraction Classification, detection
U-Net Encoder-decoder with skip connections Segmentation
ResNet Residual connections for depth Image classification
YOLO/Faster R-CNN Object detection networks Lesion detection
Transformer/ViT Attention-based architecture Multi-modal analysis
GAN (Generative) Generate synthetic images Data augmentation, super-resolution

Transfer Learning

Transfer learning leverages pre-trained networks (ImageNet) as starting points for medical imaging tasks. This approach is crucial when labeled medical data is limited. Fine-tuning adapts generic visual features to specific clinical applications, dramatically reducing training data requirements.

6.3 Clinical Applications

AI applications span the entire imaging workflow:

Detection and Diagnosis

Application Modality Performance
Lung Nodule Detection CT Sensitivity >90%, reduced false positives
Breast Cancer Screening Mammography Matches expert radiologists
Diabetic Retinopathy Fundus Photography First FDA-approved autonomous AI
Chest X-ray Triage Radiography Priority flagging for urgent cases
Brain Hemorrhage CT Seconds to alert, notification systems
Bone Fractures X-ray Reduced missed fractures

Workflow Optimization

6.4 Computer-Aided Detection (CAD)

CAD systems highlight potential abnormalities for radiologist review:

CAD Evolution

Generation Technology Characteristics
First Generation Rule-based algorithms High false positive rates
Second Generation Machine learning Better specificity, still prompts-based
Third Generation Deep learning Near-human performance
Autonomous AI Independent diagnosis No radiologist review required

Second Reader vs. Concurrent Read

CAD can function as a "second reader" (viewed after initial interpretation) or concurrently (displayed during review). Studies suggest concurrent reading may improve efficiency while second reading may improve detection rates. The optimal approach depends on the clinical setting and AI performance characteristics.

6.5 Regulatory Framework

AI medical devices face rigorous regulatory requirements:

FDA SaMD Classification

Risk Level Class Pathway Examples
Low Class I Exempt / 510(k) Workflow tools
Moderate Class II 510(k) CAD assist, measurements
High Class III PMA Autonomous diagnosis

⚠️ Regulatory Challenges

Continuously learning AI systems pose regulatory challenges: How to evaluate algorithms that change over time? FDA has proposed frameworks for "predetermined change control plans" that pre-specify allowable modifications. Post-market surveillance and real-world performance monitoring are increasingly important.

6.6 Implementation Considerations

Successful AI deployment requires careful planning:

Integration Requirements

Validation and Testing

Phase Objective Considerations
Technical Validation Verify performance claims Local data testing, edge cases
Clinical Validation Assess clinical impact Workflow integration, user acceptance
Silent Mode Testing Background operation Compare AI to clinical decisions
Prospective Monitoring Ongoing performance Drift detection, outcome tracking

6.7 Explainability and Trust

Understanding AI decisions is crucial for clinical adoption:

Explainability Methods

Radiologist-AI Collaboration

The optimal model is human-AI collaboration, not replacement. AI handles pattern recognition and routine tasks while radiologists provide clinical context, handle complex cases, and maintain final responsibility. Studies show the combination often outperforms either alone.

6.8 Future Directions

Emerging trends in medical imaging AI:

Advancing Technologies

✓ AI Implementation Success Factors

Key success factors: clear clinical use case with measurable benefit; rigorous local validation before deployment; seamless workflow integration; change management and user training; ongoing monitoring and feedback loops; transparent communication about AI capabilities and limitations.

📌 Key Takeaways

📝 Review Questions

  1. Compare CNN, U-Net, and Transformer architectures. What tasks is each best suited for in medical imaging?
  2. What is transfer learning and why is it important for medical imaging AI development?
  3. Describe the evolution of CAD systems from first generation to autonomous AI. What changed?
  4. Explain the FDA's SaMD classification framework. How does risk level affect regulatory pathway?
  5. What validation steps should be performed before deploying an AI algorithm in clinical practice?
  6. How do explainability methods like Grad-CAM help build trust in AI decisions?

Chapter 6 — Notes & References

  1. WIA Standards Public Repository (medical-imaging folder), MIT License, GitHub: WIA-Official/wia-standards-public/tree/main/medical-imaging — open standard initiative providing source code for simulator, spec, API, and ebook assets cited throughout this volume; serves as the canonical verification record for all primary-source citations made by the WIA standard committee in this chapter. Canonical ENUM tokens used in this volume include CT, MRI, PET, SPECT, MAMMOGRAPHY, ULTRASOUND, XRAY, DICOM_3_0, NIFTI, NRRD, JPEG_BASELINE, JPEG_2000_LOSSLESS, HEVC, DICOMWEB, QIDO_RS, STOW_RS, WADO_RS, PACS, VNA, RIS, HIS, FDA_510K, CE_MDR, MFDS_CLASS_2, HIPAA, GDPR, DICOM_CONFORMANCE, RUL, RML, RLL, LUL, LLL, MEDIASTINUM, HL7_FHIR_IMAGING, IHE_RAD, NEMA_MITA, DICOM_SR, RADLEX, U_NET, V_NET, RESNET, VIT, TRANSFORMER, CNN, MONAI.

6.9 Simulator ENUM Tokens and Threshold Mapping

This section aligns Chapter 6 AI and machine learning content with simulator panel 2 (Protocol) at the token level. AI architecture ENUMs, certification ENUMs, and standard ENUMs map 1:1 with the FDA Software as a Medical Device (SaMD) classification framework. ENUM tokens function as common identifiers across the FDA AI/ML Action Plan (2024), Korea MFDS AI Medical Device Approval Guidelines (2024), and EU MDR (2017/745) evaluations.

6.9.1 AI Architecture ENUM Mapping

Table 6.9.1 Medical imaging AI architecture ENUMs and Chapter 6 sections
ENUM TokenArchitectureOrigin (Year)Medical Imaging Application
CNNConvolutional Neural NetworkLeCun 1989; AlexNet 2012Classification, detection backbone
U_NETU-NetRonneberger, MICCAI 20152D medical segmentation standard
V_NETV-NetMilletari, 3DV 20163D volumetric segmentation
RESNETResNetHe et al., CVPR 2016Chest X-ray classification, residual learning
VIT / TRANSFORMERVision Transformer / TransformerDosovitskiy, ICLR 2021Pathology slides, whole-volume
GANGenerative Adversarial NetworkGoodfellow, NeurIPS 2014Synthetic data, modality translation
DIFFUSION_MODELDiffusion modelHo et al., NeurIPS 2020Low-dose CT denoising, synthesis
NNDETECTIONnnDetectionBaumgartner, MICCAI 2021Medical object detection framework
MONAIMONAI FrameworkNVIDIA/King's College, 2020PyTorch-based medical AI platform
NIFTYNETNiftyNetUCL, 2018TensorFlow-based segmentation
SEGRESNETSegResNetMyronenko, 2018 BraTSBrain tumor segmentation (winning model)

6.9.2 Certification and Regulatory ENUMs

FDA SaMD classification applies directly to medical imaging AI devices. FDA_510K represents the substantial-equivalence pathway, CE_MDR the EU Medical Device Regulation (2017/745), MFDS_CLASS_2 the Korean MFDS Class 2 (moderate risk), HIPAA the US patient information protection statute, GDPR the EU general data protection regulation, and DICOM_CONFORMANCE the DICOM Conformance Statement. The 2024 FDA AI/ML Action Plan defines Good Machine Learning Practice (GMLP) principles for medical AI: data quality, training rigor, validation, generalization, explainability, monitoring, transparency, and lifecycle management. IHE_RAD, HL7_FHIR_IMAGING, DICOM_SR, and RADLEX ENUMs support standardized AI result reporting, EHR integration, and vocabulary alignment.

6.10 Medical Imaging AI Architecture Deep Dive

6.10.1 U-Net Contracting and Expanding Paths

U-Net comprises a contracting path (encoder) of Conv-BN-ReLU blocks and an expanding path (decoder) with upsampling and skip connections. The original Ronneberger MICCAI 2015 paper achieved IoU 0.9203 on EM neural membrane segmentation, and since then it has become the de facto standard for chest X-ray lung segmentation (Dice 0.97), cardiac MRI left ventricle segmentation (Dice 0.95), and multi-organ abdominal CT segmentation. Skip connections transmit high-resolution encoder features to the decoder, preserving boundary accuracy. The 3D U-Net variant by Çiçek et al. adds z-axis convolutions optimized for volumetric segmentation, while V-Net by Milletari et al. extends the architecture with residual blocks. SegResNet, winner of BraTS 2018, is the standard for brain tumor segmentation.

6.10.2 Vision Transformer (ViT) Patch Embedding

ViT partitions input images into 16x16 patches, embeds each as tokens, and processes them through a Transformer encoder. In medical imaging, hybrid architectures dominate: SwinUNETR, MedFormer, and TransUNet combine the inductive biases of CNN with the global receptive field of attention. Recent commercial deployments employ ViT backbones in chest nodule, pneumothorax, and cardiomegaly detection achieving AUC 0.989 in multi-site validation. The MONAI Framework (NVIDIA/King's College) includes ViT backbones as standard modules.

6.10.3 Diffusion Models and Synthetic Data

Denoising Diffusion Probabilistic Models (DDPM) and latent diffusion progressively recover images from noise. In medical imaging, applications include (a) low-dose CT denoising, (b) accelerated MRI reconstruction, (c) synthetic image generation for data augmentation, and (d) modality translation (CT to MRI). Stable Diffusion Medical fine-tuned models synthesize brain MRI slices conditioned on text, mitigating training-data scarcity for rare diseases. Medical applications since the Ho et al. NeurIPS 2020 paper are surveyed in Medical Image Analysis (MedIA) 2024 vol.91.

6.10.4 Self-Supervised Learning (SSL) and Foundation Models

Because medical imaging labels are expensive, self-supervised learning is decisive. SimCLR, MoCo, DINO, and MAE (Masked Autoencoder) learn visual representations without labels. SAM Medical (Meta's Segment Anything fine-tuned), BiomedCLIP, and MedSAM emerge as medical foundation models. Regulatory frameworks now require Training Data Report, Fine-Tuning Procedure, and External Validation Result for foundation-model-based medical AI. Jun Ma et al. MedSAM (arXiv:2304.12306, 2024) has become the standard SAM variant for medical imaging.

6.10.5 Clinical Validation and External Cohorts

The gold standard for medical AI validation is multi-center, multi-national, prospective external validation. Leading mammography AI products have been validated on cohorts from Korea, the UK, the United States, Hungary, and Sweden, demonstrating generalization across ethnicities, devices, and protocols. Chest X-ray AI products often combine 700,000+ domestic studies with 95,000+ NIH ChestX-ray14 cohorts to achieve simultaneous FDA 510(k) and CE MDR approvals. Pranav Rajpurkar et al. Nature Medicine 28 (2022) and Andre Esteva et al. npj Digital Medicine 4 (2021) are standard references for large-scale medical imaging AI validation.

6.10.6 Explainability (XAI) and Clinical Trust

Grad-CAM, Integrated Gradients, and SHAP are standard techniques for visualizing model decision rationale. Explainability is the dual gatekeeper for clinical trust and regulatory approval. The 2024 FDA AI/ML Action Plan elevates GMLP Principle 8 ("clinicians must be able to understand model outputs") into a de facto XAI mandate. The MONAI framework provides standardized XAI hooks compatible with PyTorch Lightning training pipelines, enabling reproducible XAI generation across imaging AI workloads.

6.11 Extended Endnotes

  1. Olaf Ronneberger et al., "U-Net: Convolutional Networks for Biomedical Image Segmentation", MICCAI 2015 ────.
  2. Fausto Milletari et al., "V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation", 3DV 2016 ────.
  3. Özgün Çiçek et al., "3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation", MICCAI 2016 ────.
  4. MONAI Project, "MONAI Framework Documentation", NVIDIA/King's College London, monai.io, 2024 ────.
  5. NiftyNet Project, "NiftyNet: a TensorFlow-based platform for medical imaging", UCL, 2018 ────.
  6. Alexey Dosovitskiy et al., "An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale", ICLR 2021 ────.
  7. Jun Ma et al., "Segment Anything in Medical Images (MedSAM)", arXiv:2304.12306, 2024 ────.
  8. Jonathan Ho et al., "Denoising Diffusion Probabilistic Models", NeurIPS 2020 ────.
  9. FDA, "Artificial Intelligence and Machine Learning Action Plan", U.S. Food and Drug Administration, 2024 ────.
  10. Andre Esteva et al., "Deep learning-enabled medical computer vision", npj Digital Medicine 4:5, 2021 ────.
  11. Pranav Rajpurkar et al., "AI in health and medicine", Nature Medicine 28:31-38, 2022 ────.
  12. Eric Topol, "High-performance medicine: the convergence of human and artificial intelligence", Nature Medicine 25:44-56, 2019 ────.
  13. Daniel S. W. Ting et al., "AI for medical imaging goes deep", Nature Medicine 24:539-540, 2018 ────.
  14. Geert Litjens et al., "A survey on deep learning in medical image analysis", Medical Image Analysis 42:60-88, 2017 ────.