- 期刊
基於人工智慧的皮膚病理影像辨識
Image Recognition on Pathology of the Skin Based on Artificial Intelligence
摘要
目的：人工智慧快速的發展，已經吸引很多人的興趣，關注它在醫學影像問題上的應用。隨著人工智慧時代的來臨，在人工智慧中，深度學習已經變成主要的技術，它在影像處理方面能力很強。我們想藉由深度學習，建立個人化人工智慧助理，可幫助皮膚科醫師在皮膚病理方面做出正確的診斷。此外，病理科醫師在傳送病患病理診斷到健保雲端前，可利用此助理再度確認診斷。方法：在低階深度學習框架中，選擇最常用的Tensorflow框架。在高階深度學習框架（High-Level DL Frameworks）中，則選擇Keras框架。我們訓練了一個卷積神經網路模型來辨識皮膚病理。計算辨識準確率、損失函數值、k-fold cross-validation、精確率、召回率、AUC（area under the ROC curve）值、ROC曲線（Receiver Operating Characteristic curve）來偵測此卷積神經網路模型表現。結果：訓練組準確率及損失函數值分別為86.23%及0.3363。測試組準確率及損失函數值分別為84.60%及0.3957。此結果顯示個人化人工智慧助理表現良好。結論：人工智慧在皮膚病理影像辨識方面是有幫助的。
並列摘要
Objective: The rapid development of artificial intelligence(AI), has drawn much interest in its application to medical imaging. With the advent of AI, deep learning-a part of artificial intelligence, has become the dominant technique as they demonstrate considerable capabilities in image processing. Thus, by using deep learning, we want to build an AI personal assistant, which may help the dermatologist get the correct diagnosis on pathology of the skin. In addition, pathologists can use this assistant to reconfirm the diagnosis before transmitting the patient's pathological diagnosis to the National Health Insurance MediCloud. Methods: Among Low-Level Deep Learning Frameworks, we used the Tensorflow framework as it is the most actively used. Among High-Level Deep Learning Frameworks, we used the Keras framework. We trained a convolutional neural network model to recognize pathology of the skin. The accuracy, loss function value, k-fold cross-validation, precision, recall, receiver operating characteristic (ROC) curve, and area under the curve (AUC) were calculated for the performance of this convolutional neural network model. Results: The accuracy and loss function value of train dataset were 86.23% and 0.3363, respectively. The accuracy and loss function value of test dataset were 84.60% and 0.3957, respectively. The performance of the AI personal assistant is good. Conclusions: Artificial intelligence is beneficial to image recognition on pathology of the skin.