Interventional and Experimental Endoscopy (InExEn), Department of Internal Medicine 2, University Hospital Würzburg, Oberdürrbacherstr. 6, 97080, Würzburg, Germany.
Clinic for General Internal Medicine, Gastroenterology, Hepatology and Infectiology, Pneumology, Klinikum Stuttgart-Katharinenhospital, Kriegsbergstr. 60, 70174, Stuttgart, Germany.
Sci Rep. 2024 Aug 13;14(1):18825. doi: 10.1038/s41598-024-68866-x.
Standardized assessment of the gastroesophageal valve during endoscopy, attainable via the Hill classification, is important for clinical assessment and therapeutic decision making. The Hill classification is associated with the presence of hiatal hernia (HH), a common endoscopic finding connected to gastro-esophageal reflux disease. A novel efficient medical artificial intelligence (AI) training pipeline using active learning (AL) is designed. We identified 21,970 gastroscopic images as training data and used our AL to train a model for predicting the Hill classification and detecting HH. Performance of the AL and traditionally trained models were evaluated on an external expert-annotated image collection. The AL model achieved accuracy of 76%. A traditionally trained model with 125% more training data achieved 77% accuracy. Furthermore, the AL model achieved higher precision than the traditional one for rare classes, with 0.54 versus 0.39 (p < 0.05) for grade 3 and 0.72 versus 0.61 (p < 0.05) for grade 4. In detecting HH, the AL model achieved 94% accuracy, 0.72 precision and 0.74 recall. Our AL pipeline is more efficient than traditional methods in training AI for endoscopy.
通过希尔分类法对内镜下胃食管瓣进行标准化评估，对于临床评估和治疗决策至关重要。希尔分类法与食管裂孔疝（HH）的存在相关，HH 是一种常见的内镜发现，与胃食管反流病有关。设计了一种使用主动学习（AL）的新型高效医学人工智能（AI）训练管道。我们确定了 21970 张胃镜图像作为训练数据，并使用我们的 AL 训练了一个用于预测希尔分类和检测 HH 的模型。AL 模型和传统训练模型的性能在外部专家注释的图像集合上进行了评估。AL 模型的准确率为 76%。具有 125%更多训练数据的传统训练模型的准确率为 77%。此外，AL 模型在稀有类别上的精度高于传统模型，对于 3 级的精度为 0.54，对于 4 级的精度为 0.72（p<0.05）。在检测 HH 方面，AL 模型的准确率为 94%，精度为 0.72，召回率为 0.74。与传统方法相比，我们的 AL 管道在训练内镜 AI 方面更高效。
