| 摘要: | 研究目的
本研究旨在利用人工智慧深度學習技術，從術前眼部共軛斷層掃描儀(optical coherence tomography, OCT)影像中建構一個模型，預測黃斑部皺褶(epiretinal membrane, ERM)手術的術後結果。
材料與方法
本研究收錄了來自607位病患、644隻眼睛的644張術前OCT影像進行內部訓練和驗證，以及使用來自46位病患、52隻眼睛的52張術前OCT影像進行外部測試。術後一年，若snellen視力表上的視力提高了?2行，則分類為“顯著視力改善組”，若術後一年後snellen視力表的視力提高了<2行，則分類為“有限視力改善組”。我們將數據切割為80%用於訓練和20%用於驗證。使用了三個遷移式學習模型Inception-v3、ResNet-101和VGG-19進行了訓練，並使用Grad-CAM進行了熱點分析。最後，邀請了六名眼科醫生評估外部測試數據集，並將其判斷與我們的模型進行比較。
結果
在三個預訓練模型中，ResNet-101表現最佳，其Grad-CAM熱圖分析與臨床醫生的邏輯非常相似。其性能指標為召回率0.90，特異性0.90，精確度0.91，F1 分數0.90，準確度0.90 和曲線下面積(AUC) 0.97。與眼科醫師相比，這個深度學習模型的性能顯著優於一般眼科醫師及視網膜專科以外之次專科眼科醫師，略優於資淺之視網膜專科醫師。
結論
利用深度學習模型分析術前OCT 影像在預測黃斑部皺褶手術預後方面具有良好成效，這不僅有助於臨床眼科醫師更好地理解手術預後，還能協助臨床研究OCT 影像中所觀察到的顯微結構特徵。這些資訊可為臨床醫生提供更全面的患者資料，從而做出更佳的臨床決策，並更有效地制定黃斑部皺褶患者的手術計劃。
Purpose
The purpose of this study is to utilize artificial intelligence deep learning techniques to construct a model predicting the postoperative outcomes of idiopathic epiretinal membrane (ERM) surgery from preoperative optical coherence tomography (OCT) images.
Materials and methods
A total of 644 OCT images from 644 eyes of 607 subjects were utilized for internal training and validation, and 52 OCT images from 52 eyes of 46 subjects were utilized for external testing. Those with an increase of ?2 lines on the Snellen chart one year after surgery were classified as “Pronounced visual improvement group ”, while those with an increase of <2 lines on the Snellen chart one year after surgery were classified as “ Limited visual improvement group.” Data was split into 80% for training and 20% for validation. Three transfer learning models, Inception-v3, ResNet-101, and VGG-19 were trained using this data, and Grad-CAM was employed for hotspot analysis. Finally, six ophthalmologists were invited to assess the external testing dataset and compare their judgments with our model.
Results
Among the three pre-trained models, ResNet-101 performed the best, and its Grad-CAM heatmap analysis closely resembled the logic of clinical physicians. Its performance metrics were recall 0.90, specificity 0.90, precision 0.91, F1-score 0.90, accuracy 0.90, and AUROC 0.97. Compared to human classification, the performance of this deep learning model was significantly superior to general ophthalmologists and slightly better than young retina specialists.
Conclusion
Preoperative OCT image analysis using deep learning shows promise in predicting ERM surgery prognosis, aiding ophthalmologists in understanding outcomes and structural mechanisms observed in OCT images. |
