Clinical endoscopy · 2026
Impact of expert-curated video training data on computer-aided detection of sessile serrated lesions: a retrospective study in Korea.
СтатьяНазвание 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.
Han Jaehee, Oh Sang-Il, Kim Piljoo, Kim Kyung-Nam
Аннотация
BACKGROUND/AIMS: Computer-aided detection (CADe) improves adenoma detection; however, its performance for sessile serrated lesion (SSL) detection remains inconsistent. We hypothesized that models trained on expert-curated, histopathologically-confirmed datasets would improve SSL detection compared to models trained on public datasets without verified labels.
METHODS: Two CADe models with identical Visual Geometry Group 16 architectures were trained using different datasets: Model A on public datasets (~54,000 frames) and model B on a histopathologically-confirmed private dataset (~120,000 frames) derived from examinations performed by endoscopists with adenoma detection rates of >35%. Validation was conducted using 31 independent colonoscopy videos obtained from another endoscopist. Diagnostic performance was evaluated using event- and frame-based analyses.
RESULTS: Model B demonstrated a significantly higher event-based sensitivity than model A (99.7% vs. 39.9%, p<0.001). The detection of SSLs was markedly improved with model B. Frame-based sensitivity and F1-score were also higher for model B. Although model B generated more false positives per video (2.55 vs. 0.16, p<0.001), these alerts were brief and unlikely to meaningfully interfere with the endoscopic workflow.
CONCLUSIONS: CADe models trained on expert-curated, histopathologically-confirmed datasets showed improved detection of neoplastic colorectal lesions, including SSLs, compared to models trained on public datasets. These findings highlight the importance of clinically curated datasets for optimizing the CADe performance for subtle colorectal lesions.
METHODS: Two CADe models with identical Visual Geometry Group 16 architectures were trained using different datasets: Model A on public datasets (~54,000 frames) and model B on a histopathologically-confirmed private dataset (~120,000 frames) derived from examinations performed by endoscopists with adenoma detection rates of >35%. Validation was conducted using 31 independent colonoscopy videos obtained from another endoscopist. Diagnostic performance was evaluated using event- and frame-based analyses.
RESULTS: Model B demonstrated a significantly higher event-based sensitivity than model A (99.7% vs. 39.9%, p<0.001). The detection of SSLs was markedly improved with model B. Frame-based sensitivity and F1-score were also higher for model B. Although model B generated more false positives per video (2.55 vs. 0.16, p<0.001), these alerts were brief and unlikely to meaningfully interfere with the endoscopic workflow.
CONCLUSIONS: CADe models trained on expert-curated, histopathologically-confirmed datasets showed improved detection of neoplastic colorectal lesions, including SSLs, compared to models trained on public datasets. These findings highlight the importance of clinically curated datasets for optimizing the CADe performance for subtle colorectal lesions.