JRC2026 symposium 1 KATAYAMA

JRC2026 Joint Symposium 1 2026/04/17 The Future of Clinical Specialties with LLMs and AI: A Radiological Technologist's Perspective. The presentation slides can be accessed here. ⇩ Yutaka Katayama1,2 1. Department of Radiology, Osaka Metropolitan University Hospital 2. Advanced Imaging Technology Laboratory Division of Health 2. Sciences, The University of Osaka Graduate School of Medicine Unattributed images are either original experimental data or AI-derived.

Disclosure of Conflict of Interest (COI) We have nothing to declare for this study. The 82nd Annual Meeting of JSRT Japanese Society of Radiological Technology

The State of Clinical AI 2026 • Rapid Integration: 1,200+ FDA-cleared tools now embedded in daily care1. • The Reality Check: Bridging the gap between Lab benchmarks and Clinical uncertainty. • From Tool to Teammate: Shifting focus from Replacing humans to Augmenting judgment. • Outcome-Driven: Moving beyond technical hype toward real-world patient benefits. 1. https://medicine.stanford.edu/news/current-news/standard-news/clinical-ai-has-boomed.html

Contents ① The Tide of Change: Transformations Brought by Generative AI ② Paradigm Shift: The Future of Radiology with Generative AI ③ Future Outlook: The New Value of Radiological Technologists in the Era of Generative AI

Contents ① The Tide of Change: Transformations Brought by Generative AI ② Paradigm Shift: The Future of Radiology with Generative AI ③ Future Outlook: The New Value of Radiological Technologists in the Era of Generative AI

① Trends in Image Processing • Direct Impact of AI on Radiological Technologists ⇒ The Emergence of Deep Learning Reconstruction • Core function • Suppresses noise and other image-degrading components while preserving diagnostically relevant structures. • Practical impact • Enables a better balance among safety，efficiency， and image quality in daily imaging practice. • Clinical outcomes • Supports lower-dose imaging in computed tomography (CT). • Supports shorter acquisition in magnetic resonance imaging (MRI). • Improves overall image quality across modalities.

DLR via AI-driven Restoration Original Processed Image

Data-Driven 3D Reconstruction • [Left] Model generating back view from frontal image • [Right] Model generating cross-section from bi-plane X-rays ⇒ In data-rich fields, 3D reconstruction from 2D images is widely explored. 2. Saito, Shunsuke, et al. "Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020. 3. Ying, Xingde, et al. "X2CT-GAN: reconstructing CT from biplanar X-rays with generative adversarial networks." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2019.

Data-Driven 3D Reconstruction DRR (Frontal) DRR (Lateral） Image reconstructed via self-supervised learning

Geometry-Driven 3D Reconstruction Reconstruction Algorithm Output: 3D Shape Viewpoint 2 Viewpoint 1 Input: 2D Image • 3D reconstruction from dual views via Computer Vision (CV) • No training required (CV-based approach) • This work is currently patent pending.

Artifact Detection via Vision Transformer • Anomaly detection model trained solely on metal-free normal images • Reconstruction errors from metal-containing inputs reveal streak artifacts and dark bands • Ensemble learning quantifies/visualizes detection reliability

Segmentation (Nodule) Automated model generation using JSRT Nodule154images & CLNDAT_EN.txt 1. Mask generation via Generative AI • Masks unavailable; coordinates extracted from CLNDAT_EN.txt used • Nodule shape inference via Segment Anything Model (SAM) using images/coordinates • SAM failures replaced by circles based on size data ⇒ High-quality training masks generated via Generative AI ⇒ Human-out-of-the-Loop 2. Segmentation • Automatic data split: Training (80%) / Validation (20%) • Training weighted to prioritize nodules against background dominance ⇒ Segmentation performed via SegFormer

Segmentation (AI Challenge in JRC2026) JRC2026 AI Challenge solution (Python) • Segmentation model based on ResUNet (U-Net + Residual Blocks) • Automatic data split: Training (80%) / Validation (20%) ⇒ Model achieved Average Dice Coefficient of 0.96

② Trends in LLM Potential • Ensuring Clinical Accuracy • Advances in training and prompt engineering to mitigate hallucinations. • Strict verification of LLM outputs remains mandatory for clinical use. • Our Implementation: Clinical Agents 1. Vision-LLM Agents: For image-anchored reasoning. 2. Interview Agents: For structured pre-examination data collection.

Local VLMs Are All You Need Input Image Output Result • Vision-Language Model (VLM): AI model processing interplay between visual and linguistic information • LLaVA implemented in Mac virtual environment • Extensive image processing/analysis enabled via OpenCV • Configured to detect metal artifacts and suggest countermeasures • Local execution addresses privacy concerns

Local LLMs Are All You Need • Developed offline (Local) LLM using Python • Adopted Japanese-specific LLM (Mistral & NVIDIA) due to hardware constraints (CyberAgent model unusable) • Hybrid system (Rule-based + LLM) developed to compensate for standalone LLM limitations

Contents ① The Tide of Change: Transformations Brought by Generative AI ② Paradigm Shift: The Future of Radiology with Generative AI ③ Future Outlook: The New Value of Radiological Technologists in the Era of Generative AI

Future of Generative AI in Radiology Outcome 1: Advancement & Efficiency in Diagnosis 1. Image Optimization & Quality Control via AI • Imaging optimization & noise-reduced reconstruction ⇒ High-speed imaging and reduced examination time ⇒ High-quality diagnostic images ⇒ Enhanced information via 2D/3D conversion 2. Abnormality Detection & Verification • AI suggests abnormal region candidates; enables efficient acquisition-time oversight ⇒ Automated verification ⇒ of diagnostic information capture

Future of Generative AI in Radiology Outcome 2: Workload Reduction 1. Automation & Optimization of Imaging Process • Auto-parameterization & positioning ⇒ Time reallocation to patient care ⇒ Supports complex procedures ⇒ Enables increased examination throughput 2. Safety & Dose Optimization • Patient-specific dose optimization ⇒ High-quality diagnostics at lower dose ⇒ AI-assisted safety assurance

Future of Generative AI in Radiology Outcome 3: Shift to Preventive Medicine 1. Preventive Medicine4 • Early detection and continuous monitoring automated; Shift from curative to preventive medicine facilitated 4. Ohta, Yukino, et al. "Development of AI model for dual detection of low bone mineral density in the femoral neck and lumbar vertebrae using chest radiographs." Journal of Clinical Densitometry 28.4 (2025): 101604.

Future of Generative AI in Radiology Issue 1: Ethical & Social Issues 1. Hallucination & Information Accuracy • Potential for AI-generated misinformation; Rigorous human verification required 2. Data Privacy & Security • Handling sensitive patient data requires anonymization, robust security, and strict privacy protection 3. The Black Box Problem • Transparency in AI diagnosis/judgment essential; Understanding rationale needed for reliability ⇒ Potential solution via Explainable AI, though cost remains an issue

Future of Generative AI in Radiology Issue 2: Legal & Social Issues 1. Locus of Responsibility • AI medical liability needs clarification. ⇒ Currently, physicians are the primary agents ⇒ Physicians explicitly bear final responsibility 2. Ensuring Fairness • Training data bias may cause diagnostic disparities ⇒ Ensuring fairness is a key challenge • Generative AI holds potential to revolutionize radiology; Implementation requires technical progress, deep ethical/social discussion, careful consideration, and regulatory frameworks

Contents ① The Tide of Change: Transformations Brought by Generative AI ② Paradigm Shift: The Future of Radiology with Generative AI ③ Future Outlook: The New Value of Radiological Technologists in the Era of Generative AI

AI Readiness Gaps: Skills & Education • Shortage of AI-Literate Personnel • Few pros can evaluate and utilize AI. • Education: Avoid blind dependence or distrust. • Lack of Established Training Systems • Rapid tech evolution outpaces safety standards and guidelines. • Urgent need to establish a framework defining who, what, and how to teach.

Neologisms from LLM Misreading • LLMs generated non-existent words due to data reading errors, and the phenomenon where this was cited in actual papers was confirmed. • The non-existent term "vegetative electron microscopy" was accidentally created5) by a layout error during digitization. It was incorporated into AI training data and cited in subsequent papers6). • Recognized as a new challenge for ensuring information reliability as AI-assisted research and writing grows. 5) 6) 5. Strange, R. E. "Cell wall lysis and the release of peptides in Bacillus species." Bacteriological Reviews 23.1 (1959): 1-7. 6. Alishiri, Mostafa, et al. "Study of CNT@ Fe3O4 effects on Aeromonas hydrophila and Yersinia ruckeri bacteria isolated from fish." Journal of Fisheries 72.1 (2019).

Cheating LLMs (?) • In the arXiv paper shown here, a message addressed to an LLM was visible near the abstract in version 3, although it was removed in the latest version, v4. Displayed in the HTML version of v3. 7. Ge, H., Rudzicz, F., & Zhu, Z. (2024). Understanding Language Model Circuits through Knowledge Editing. arXiv preprint arXiv:2406.17241.

Conclusion: New Value in the AI Era • AI can reduce routine and error-prone tasks. • This allows technologists to focus on safer and more reliable work. 1. Risk Management • Verify AI outputs and avoid unsafe use of uncertain information. 2. Clinical Judgment • Focus on technically demanding decisions in complex cases. 3. Team Coordination • Connect AI use with safe workflows and clinical practice.

Acknowledgments • I would like to express my deepest gratitude for this opportunity to: • Assoc. Shohei Hanaoka Dept. of Computational Diagnostic Radiology & Preventive Medicine, The Univ. of Tokyo Hospital • Assoc. Prof. Rie Tanaka College of Medical, Pharmaceutical & Health Sciences, Kanazawa Univ.