基于计算机视觉的人工智能系统（FenoParasite）用于粪便样本中蛔虫和蓝氏贾第鞭毛虫显微镜检测的开发与验证
Development and Validation of a Computer Vision-Based Artificial Intelligence System (FenoParasite) for the Microscopic Detection of Ascaris lumbricoides and Giardia lamblia in Stool Samples
在资源有限地区，肠道寄生虫诊断依赖专业人员显微镜检查，灵敏度和可及性都不稳定。本研究开发并验证了一款名为FenoParasite的人工智能软件（基于微软Azure自定义视觉），用于在粪便显微镜图像中识别蓝氏贾第鞭毛虫的包囊和滋养体以及蛔虫卵。以两位专家阅片共识为参照标准，在159例内部验证集和31例前瞻性初步验证集中评估，模型内部验证对两种寄生虫的灵敏度达100%，特异度分别为98.9%和100%。但前瞻性验证中阳性标本数很少（分别为4例和13例），可信区间很宽，且参照标准是人工显微镜而非分子检测，因此结果只是与专家阅片的一致性，尚属初步，不能作为绝对诊断准确性。临床应用前还需多中心、以分子检测为参照的验证。
为什么推荐给您：AI辅助寄生虫镜检的新工具，但仅为诊断准确性初步验证，样本量小、无分子参照，属值得关注的早期研究。
不需要生物学背景，多打比方
正在获取全文并生成讲解（拿不到全文就依据摘要），大约需要 30–60 秒…
已等待 0 秒
这篇还没有动画
动画会把研究的流程、作用机制和关键结果一步一步演示出来，每一步都标明出自原文哪里。制作大约需要 30–60 秒。
摘要Abstract
BACKGROUND: Intestinal parasitic infections caused by Giardia lamblia and Ascaris lumbricoides remain a public-health problem in resource-limited settings, where conventional microscopy depends on expert personnel and shows variable sensitivity. Artificial intelligence (AI) could standardize and decentralize diagnosis. This study aimed to design and validate FenoParasite, an AI object-detection software (Azure Custom Vision) for the microscopic identification of G. lamblia cysts and trophozoites and A. lumbricoides eggs in stool samples.
METHODS: Diagnostic-accuracy study. The reference standard was consensus microscopy by two expert readers, with a third reader for discordant cases. An internal validation set (n = 159) and a prospective pilot validation set (n = 31) were analyzed. Sensitivity, specificity, positive predictive value, negative predictive value, accuracy (95% Clopper-Pearson CI), AUC-ROC (bootstrap), and Cohen's kappa were computed; the confidence threshold was 90%.
RESULTS: The model achieved a mAP@0.30 of 98.3% (precision 98.2%; recall 96.4%). In internal validation, G. lamblia showed 100% sensitivity, 98.9% specificity, 99.4% accuracy, and AUC 0.98; A. lumbricoides reached 100% across all indices (AUC 1.00). In the prospective pilot validation, sensitivity and specificity were 100% for both taxa; however, the confidence intervals were very wide (lower sensitivity bounds of 39.8% and 75.3%, based on only 4 and 13 reference-positive specimens, respectively), and these estimates are correspondingly imprecise.
CONCLUSIONS: FenoParasite showed high agreement with consensus expert microscopy for both a protozoan and a helminth. Because the reference standard was expert microscopy rather than a molecular assay, the reported performance estimates quantify agreement with microscopic reading rather than absolute diagnostic accuracy. These findings remain preliminary. Multicenter validation against molecular reference standards is required before clinical implementation can be considered.