| 摘要: | 背景 息肉尺寸為決定大腸鏡追蹤間隔之關鍵因素之一。本研究提出一種無需參考物即可於大腸鏡檢查過程中量測息肉尺寸大小的人工智慧模型。
材料與方法 本研究建立之息肉尺寸估測迴歸模型，係以兩個獨立之 SegFormer 模型輸出為基礎，其中一者用於息肉分割；另一者用於深度估測。模型初期以仿體在模擬大腸模型進行訓練，隨後透過 1,304 張臨床影像進行遷移式學習。測試階段使用獨立於訓練集之外的52個息肉，共178張影像。以套圈法作為尺寸比較之真實值，採用Olympus 290系列大腸鏡（視野角170°）進行拍攝。息肉依尺寸分為三組：≦5mm、5–10mm 與 ≧10mm。統計分析包括誤差率分析、召回率、精確度、Bland-Altman分析圖、配對 t 檢定及 Cohen′s kappa，以評估套圈法與AI 模型之間在量化測量與分類一致性上的表現。
結果 於測試模型中，AI 模型在三種息肉尺寸組別之誤差率分別為 10.74%、12.36% 與 9.89%，平均誤差率為 11.47%。整體召回率為 0.846；三組息肉尺寸的精確度分別為 0.870、0.911 與 0.857，平均精確度為 0.879；整體 F1 分數為 0.861。Cohen′s kappa 值為 0.792，顯示兩種方法間具高度一致性。Bland-Altman 分析顯示兩方法間之平均偏差為 -0.03 mm，一致性界限介於 -1.654mm 至 1.596mm。
結論 本研究所建立之人工智慧模型可於結腸鏡檢查中，在無需參考物的情況下，準確量測大腸息肉尺寸，具臨床應用潛力與發展前景。
;Background Polyp size is one of the key factors in determining colonoscopy surveillance intervals. We present an artificial intelligence model for colon polyp size measurement that does not require a reference object during a colonoscopy.
Materials and Methods The regression model for polyp size estimation was developed using the outputs from two independent SegFormer models, one for polyp segmentation and the other for depth estimation. Initially, colonoscopic images of polyp phantoms were used to build the model, followed by transfer learning on 1,304 real-world images. For model testing, 178 polyp images from 52 polyps, independent of the training set, were evaluated. A snare was used as the ground truth for size comparison with the AI-based model. Olympus 290 series colonoscope with field of view (FOV) of 170 angle was used in this study. Polyps were categorized into three size groups: ≦5 mm, 5–10 mm, and ≧10 mm. Statistical analysis include error rate analysis, recall, precision, Bland–Altman plot, paired t-test, and Cohen′s kappa to evaluate both the quantitative agreement and categorical consistency between the snare method and AI-based model.
Results The error rates for the snare method and the AI-based model across the three polyp size groups in testing model were 10.74%, 12.36%, and 9.89%, respectively, with an average error rate of 11.47%. The overall recall was 0.846, and precision rates for the three size groups were 0.870, 0.911, and 0.857, resulting in an average precision of 0.879. The overall F1 score was 0.861. Cohen′s kappa value between the two methods across the three groups was 0.792. Bland-Altman analysis showed a mean bias of -0.03 mm between the two methods, with limits of agreement, from -1.654 mm to 1.596 mm.
Conclusion Our AI-based model shows promise as an accurate tool for colorectal polyp size measurement without the need for a reference object during screening colonoscopy. |
