Diagnostic imaging artificial intelligence (AI) is being deployed for reader assistance, triage, quantification, and reporting. Yet strong internal-test performance often fails to yield clinical impact, because errors stem from data issues, evaluation design, and workflow/operational mismatches—not only model architecture. Based on an invited lecture, this article summarizes practical pitfalls and countermeasures for building clinically effective imaging AI through close physician–engineer collaboration. It addresses (i) failure modes such as shortcut learning, domain shift across sites/devices, and hidden stratification; (ii) evaluation design with leakage prevention and external holdout testing; (iii) assessment beyond area under the curve, emphasizing probability calibration and decision curve analysis (DCA); (iv) safety-oriented subgroup analysis with attention to worst-group performance; and (v) deployment practices including silent-run validation, drift monitoring, and change control. These elements are consolidated into an actionable checklist and a collaboration loop to support development, validation, and maintenance under real-world conditions. While the ideas generalize broadly, the discussion mainly considers binary risk prediction where output probabilities guide threshold-based decisions (e.g., triage or additional testing). This is a practice-oriented implementation perspective, not a comprehensive literature review.

This article reports on the clinical practice of radiation therapy for dogs and cats at the Koganei Animal Medical Emergency Center, Tokyo University of Agriculture and Technology. The treatment system using a high-precision linear accelerator, case distribution, characteristics of treated sites, irradiation procedures under general anesthesia, and quality assurance management are outlined. Challenges specific to veterinary medicine, including anesthesia management and limited human resources, are also addressed. Furthermore, the social significance of radiation therapy as a treatment option aligned with owners’ preferences is discussed.

This study aims to develop and evaluate an object detection model for identifying computed tomography images in real-world environments, which is a core technology for a subjective computer-aided detection system. The ultimate goal of this system that integrates mixed reality devices with computer-aided detection to reduce disparities in medical care caused by variations in image interpretation skills. The proposed system detects medical images in real space, extracts regions of interest, and analyzes them in real time. In this research, we developed a prototype system that detects and crops computed tomography images within real-world environments. YOLOv8 was employed as the object detection algorithm, and nine models were trained and evaluated. Among them, the best-performing model, referred to as model_A,1.0, was trained using a mixup data augmentation technique with backgrounds generated by generative artificial intelligence. This model achieved the highest accuracy, with average precision₅₀ = 0.992 and average precision _50-95 = 0.883. These results demonstrate the potential of applying mixed reality technology for real-time, subjective computer-aided detection and highlight the feasibility of accurate medical image detection in real environments. This study serves as an early validation of the concept, suggesting that mixed reality-based computer-aided detection systems can contribute to improving diagnostic consistency and supporting clinicians in interpreting medical images.