[[alternative]]Establish a Medical Image Analysis and Identification Platform - a Case Study of NIH Chest X-ray Dataset

[[abstract]]隨著深度學習技術日新月異，大數據分析已成為各個領域的重要趨勢，近幾年，更是在醫療資訊領域有輔助的驚人效果。其中在醫療影像方面，影像解析度越高資訊越多，從中可以觀察到症狀的特徵，相對的醫生需要花費更多時間進行影像標記與做醫療診斷。從大數據中發展出電腦輔助診斷(Computer-Aided Diagnosis, CAD)以協助醫生診斷，減少誤診率的發生並透過機器學習(Machine Learning, ML)快速發展，讓其不斷改善性能與提升辨識效果，為醫生在診斷上提供足夠的支持。本研究建構模型訓練和辨識介面的使用平台，研究並簡化模型訓練撰寫程式步驟，呈現簡單又容易使用的使用者介面供使用者便於資料集管理、模型管理、訓練模型及模型辨識的應用。本研究以已開發完全的Keras做為工具建立平台系統架構，並使用Keras內建的模型CNN架構來訓練模型，將所有的訓練資訊及模型性能結果評估儲存於資料庫，以便於日後做管理與模型訓練參考使用。平台主要功能分為兩部分，第一部分資料集和模型管理，便於使用者使用及確保使用者的權益以及是否給予其他使用者做使用及參考；第二部分模型訓練及辨識，只通過參數設定的方式完成模型訓練以及為模型提供單張或批量影像辨識。本研究採用NIH公開胸部X光影像資料作為研究數據，並於結果的部分表明本平台的模型訓練架構可以達到很好的分類能力，媲美於先前的研究結果。[[abstract]]With the rapid development of deep learning technology, big data analysis has become an important trend in various fields. In recent years, it has achieved amazing results in the field of medical information. In the field of medical imaging, the higher the image resolution, the more information about the symptom features in the image, and the radiologist needs to spend more time on image labeling and diagnosis. Computer-aided diagnosis has been proven to help doctors make diagnoses through machine learning, reduce misdiagnosis, and rapid development, which can continuously improve performance and improve recognition, thereby providing doctors with full diagnostic support. This study built a platform of Image model training and identification and eliminate the trouble of users writing their own programs, and presents a simple and easy-to-use user interface for users to facilitate data set management, the steps of model management, training model, and model identification.This study uses Keras as a tool to build the platform system architecture. The main functions of the platform are divided into two parts. The first part of the data set and model management is convenient for users to use and ensure the rights of users to decide whether to give other users to use and reference. The second part is model training and identification, user cannot parameters setting to complete model training, user can use model to single or batch image recognition.This study uses NIH public chest X-ray data as research data to prove of concept, and the results partially show that the platform's model training structure can achieve good classification capabilities, comparable to previous research results