- 學位論文
深度學習訓練效率的模糊驗證模式-以田口優化的卷積神經網路在辨識肺癌醫學影像為例
Fuzzy Verification Model of the Training Efficiency of Deep Learning: An Example of Optimized Convolutional Neural Network with Taguchi Method in Recognizing Medical Image of Lung Cancer
摘要
隨著人工智慧技術的突破性發展，LeNet-5卷積神經網路近年已被各產業廣泛用來解決產業問題與提升作業效率及準確率，然而，卷積神經網路必須藉由大量影像資料來訓練所考量的因子及其參數值，因而多少會產生不穩定的預測與分類結果，且實驗次數隨著所考量的因子及水準參數值的增加而據增，亦增加成本及耗費時間。而田口方法僅需以較少的實驗次數來找出最佳化的參數組合，因此，利用田口方法來優化卷積神經網路的模式被提出，並應用於肺癌的醫學影像進行良性和惡性腫瘤的辨識與分類，且透過混淆矩陣計算準確率指標，做為評估與判斷模型在辨識分類的效能之優劣程度。但值得注意的是，準確率指標會隨著模型每次學習訓練的優劣程度會有不同的數值，以至於準確率指標為一個不確定性的變數，即便以指標的平均數做為判斷標準，亦容易受到極端值的影響而存在著誤判的可能性。有鑑於此，根據LeNet-5卷積神經網路的參數優化前後的準確率為例來定義一個辨識效能指標IACC，並推導IACC的100(1-α)%信賴區間，而在考量準確率為一個不確定性的變數的情況下，IACC的三角形態模糊數進一步被提出，並用於發展IACC的模糊統計檢定模式，以更可靠地確認田口方法優化參數的卷積神經網路模型在辨識肺癌醫學影像上有較佳的辨識效能。
並列摘要
With breakthrough developments in artificial intelligence technology, LeNet-5 convolutional neural networks have become widely applied to solve problems and enhance operation efficiency and accuracy in various industries. However, convolutional neural networks require substantial quantities of image data to train for the factors and parameter values that are considered, which more or less generates unstable prediction and classification results. Moreover, the number of experiments increases steeply with the considered factors and parameter values, as does the costs and time consumed. The Taguchi method only requires a smaller number of experiments to obtain the optimal parameter combination. We therefore employed the Taguchi method to optimize our convolutional neural network model and applied it to recognize lung cancer medical images and classify them into benign and malignant tumors. We calculated the accuracy index using a confusion matrix to gauge the classification performance of the model. It is worth noting that the accuracy index changes with the quality of model training and is therefore an uncertain variable. Even with the index mean as a judgment standard, extreme values can easily affect the judgment results. In view of this, we defined a recognition performance index IACC based on the accuracy rates before and after the parameter optimization performed by the LeNet-5 convolutional neural network and derived the 100(1-α)% confidence interval of IACC. Considering the fact that accuracy is an uncertain variable, we further defined the triangular fuzzy number of IACC and developed a fuzzy statistical test for IACC to more reliably confirm that convolutional neural network with parameters optimized using the Taguchi method has better performance in lung cancer medical image recognition.
參考文獻
延伸閱讀
- 王志明（2019）。在多GPU架構下以深度學習方法篩檢肺腫瘤及確認其良惡性〔碩士論文，國立臺灣大學〕。華藝線上圖書館。https://doi.org/10.6342/NTU201904302
- 鄧煒霖（2019）。基於深度學習技術之間質性肺病紋理分類〔碩士論文，國立屏東科技大學〕。華藝線上圖書館。https://doi.org/10.6346/NPUST201900185
- 洪聖傑（2021）。應用深度學習於肺結節CT影像醫療診斷報告自動化生成技術〔碩士論文，中山醫學大學〕。華藝線上圖書館。https://doi.org/10.6834/csmu202100072
- 陳稜鎔（2018）。電腦斷層掃描肺腫瘤良惡性判別之深度學習影像特徵擷取〔碩士論文，國立臺灣大學〕。華藝線上圖書館。https://doi.org/10.6342/NTU201804220
- 林韻孜（2017）。Deep Learning Study of CAE with the Integration of Taguchi Method and Neural Network〔碩士論文，中原大學〕。華藝線上圖書館。https://doi.org/10.6840/cycu201700874