AI眼疾辨識導入attention機制下實作出優越模型並分析商業模式

Abstract

本論文研究目的是建立AI眼疾辨識系統來輔助眼科醫生診斷，達到早期發現與早期治療，從而改善臨床結果。本論文採用來自於Kaggle之OCT影像數據集，此影像數據集由知名醫院之專業人員嚴格審核與標註，數據共有4種類別。AI辦識眼疾辨識系統之建立，首先利用轉移學習(Transfer Learning)以及注意力機制(Attention Mechanism)，其中轉移學習使用VGG16模型，注意力機制分成1path二個模組、2path二個模組，共四種模組。其次使用OCT影像之數據集來訓練神經網絡。最後對四種模組個別進行參數調整，並找出各模型最佳參數與比較四個模型的優略。最後，篩選結果是1pathb(Conv：1path_1x3 , 3x1)與2patha(Conv：2path_1x1,1x1_1x3,3x1_add)模型之Dropout_Rate設在20%，Dense_cells數目設定512、1024為佳。實驗結果Test_acc目前達92%-93%左右，稍低於優秀專業眼科醫生93.3%判斷水準[29]，但OCT醫學影像判讀速度遠遠超過眼科醫生，表示AI眼疾辨識系統存在臨床應用之價值。AI眼疾辨識輔助診斷系統之商業化要成功，應具備：(1)AI眼疾辨識輔助診斷之市場具吸引力(2)極高品質的OCT醫學影像之標註模式，(3)能夠研發優越AI辨識模型之研發精兵組織，(4)以雲端服務為主力產品之商業模式，(5)其他層面配合。前五項紮實做好就可創造公司商業價值與客戶使用價值。再透過一系列商業競爭策略之分析：(1)價值網分析，(2)GE矩陣分析，(3)SWOT分析，(4)BCG矩陣分析，(5)公司演化與競爭策略矩陣分析，來檢視Ω公司之商業模式的可行性與獲取策略方針。

The research purpose of this paper is to establish an Artificial Intelligence(AI) of eye-related diseases identification system to assist ophthalmologists in diagnosis, so as to achieve early detection and treatment, thus improving clinical results.In this paper, Optical coherence tomography (OCT) image data set from Kaggle was used. This image data set was strictly checked and marked by professionals from well-known hospitals and there are 4 types of data. The establishment of this AI eye disease identification system was first employed Transfer Learning and Attention Mechanism, in which the VGG16 model was used for transfer learning, and the Attention Mechanism was divided into two modules of 1path and two modules of 2path, totaling four modules. Moreover, the data set of OCT images was used to train the neural network. Finally, the parameters of the four modules are adjusted individually, and the best parameters of each model was found. The advantages and disadvantages of the four models were also compared. In conclusion, 1pathb (Conv:1path_1x3, 3x1) and 2patha (Conv:2path_1x1, 1x1_1x3, 3x1_add) models were selected, and the accuracy of Dropout_Rate setting 20% and the number of Dense_cells setting 512 and 1024 is better. The experimental results of Test accuracy(Test_acc) are currently about 92%-93%, which is slightly lower than the diagnostic accuracy of 94.5% of professional ophthalmologists. However, OCT medical image diagnosis is much faster than ophthalmologists, indicating that the AI of eye-related diseases identification system has clinical application value. To succeed in the business model of the AI eye-related diseases recognition assistant diagnosis system, the necessary conditions are: (1) market appeal of AI eye disease recognition and auxiliary diagnosis (2) very high quality OCT medical image mark mode, (3) to organize An excellent R&D team to develop AI models, (4) cloud services are the main products, and (5) other support. The top five can be achieved to create company value and customer value.Through a series of business competition strategy analysis: (1) network value analysis, (2) GE matrix analysis, (3) SWOT analysis, (4) BCG matrix analysis, (5) company evolution and competition strategy matrix analysis, Verify the feasibility of the business model and provide a clear strategic direction.