Digestive endoscopy : official journal of the Japan Gastroenterological Endoscopy Society · 2026
Development and Crossover Evaluation of an Artificial Intelligence-Assisted System for Solid Pancreatic Lesion Detection and Pancreatic Parenchyma Recognition in Endoscopic Ultrasonography (With Video).
СтатьяМногоцентровое исследованиеРандомизированное исследованиеVideo-Audio MediaНазвание и аннотация: Перевод в очереди
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Fujisawa Toshio, Kuwahara Takamichi, Sato Tatsuya, Takasaki Yusuke, Okuno Nozomi, Hakuta Ryunosuke, Tomishima Ko, Haba Shin, Takahara Naminatsu, Ito Yukiko, Nakai Yousuke, Hara Kazuo, Isayama Hiroyuki
Аннотация
BACKGROUND AND STUDY AIMS: Pancreatobiliary endoscopic ultrasonography (EUS) is technically demanding, and supervised training opportunities are limited. We developed an artificial intelligence (AI) overlay system for detecting solid pancreatic lesions (SPL) and recognizing pancreatic parenchyma (PP) and evaluated its effect on reader performance.
PATIENTS AND METHODS: Across six centers, two deep learning-based models were trained using expert-annotated EUS frames. We then conducted a randomized, two-sequence, two-period crossover reader study in which eight endosonographers (five novices and three experts) interpreted image sets with and without AI assistance. The primary endpoint was superiority of sensitivity for SPL detection among novices; key secondary endpoints included specificity and PP recognition.
RESULTS: From 118 patients, 120 SPL-positive/negative image sets and 160 PP-positive/negative image sets were constructed. Among novices, AI assistance improved SPL detection sensitivity (88.7% vs. 76.8%, p < 0.001) and accuracy (86.4% vs. 78.7%), while specificity met the predefined noninferiority criterion (84.2% vs. 80.5%, p < 0.001). For PP recognition, sensitivity increased numerically (86.3% vs. 83.3%) but did not meet the predefined superiority criterion (p = 0.095); specificity met the noninferiority criterion (87.8% vs. 81.0%), and accuracy increased from 82.1% to 87.0%. Among experts, sensitivity was maintained for both tasks, whereas specificity increased with AI assistance.
CONCLUSIONS: AI assistance improved SPL detection among novice endosonographers. For PP recognition, sensitivity increased without reaching statistical superiority, whereas specificity met the predefined noninferiority criterion. These findings support a potential adjunctive role for AI in EUS interpretation.
PATIENTS AND METHODS: Across six centers, two deep learning-based models were trained using expert-annotated EUS frames. We then conducted a randomized, two-sequence, two-period crossover reader study in which eight endosonographers (five novices and three experts) interpreted image sets with and without AI assistance. The primary endpoint was superiority of sensitivity for SPL detection among novices; key secondary endpoints included specificity and PP recognition.
RESULTS: From 118 patients, 120 SPL-positive/negative image sets and 160 PP-positive/negative image sets were constructed. Among novices, AI assistance improved SPL detection sensitivity (88.7% vs. 76.8%, p < 0.001) and accuracy (86.4% vs. 78.7%), while specificity met the predefined noninferiority criterion (84.2% vs. 80.5%, p < 0.001). For PP recognition, sensitivity increased numerically (86.3% vs. 83.3%) but did not meet the predefined superiority criterion (p = 0.095); specificity met the noninferiority criterion (87.8% vs. 81.0%), and accuracy increased from 82.1% to 87.0%. Among experts, sensitivity was maintained for both tasks, whereas specificity increased with AI assistance.
CONCLUSIONS: AI assistance improved SPL detection among novice endosonographers. For PP recognition, sensitivity increased without reaching statistical superiority, whereas specificity met the predefined noninferiority criterion. These findings support a potential adjunctive role for AI in EUS interpretation.