- 學位論文
以卷積神經網路應用於胸腔X光造影之影像品質合格辨識
Image Qualified Recognition for Chest X-Ray Using CNN Model
摘要
在肺部的檢查中，胸腔X光影像是最簡單的一種檢查，對肺部的狀況可以有概括性及初步的臆斷。同時，輻射劑量最低、人體傷害也最低，對於臨床病情評估十分重要。為了達到診斷的目的，影像的品質就很重要，現在往往都是依靠放射師在拍攝完後辨別影像合格與否，然而不同的放射師在合格標準上也會有些許的不同，難以達成影像的一致性。 近年來AI發展迅速，也有許多深度學習的模型被應用在醫療影像上，若能利用科技輔助，辨識胸腔X光影像是否合格，便能增加影像的品質，同時提高醫護人員的工作效率及品質。本研究將以物件定位模型Faster-RCNN與圖像分類模型ResNet作為基礎，辨別胸腔X光影像是否合格，為了達成此目的，本研究先是利用Faster-RCNN將影像中感興區框選出來，在利用ResNet分別辨別胸腔X光影像中左、右肩胛骨、頸椎與肋膈角共四個部位是否合格。 在物件定位模型中，本研究使用IoU及Y_IoU，作為目標位置定位框選準確性的依據，而在圖像分類模型則是使用準確率、靈敏度及特異度判別各部位合格標準辨識的評估指標。
並列摘要
In the examination of the lungs, chest X-ray imaging is the simplest kind of examination, which can make a general and preliminary assumption about the condition of the lungs. At the same time, the radiation dose is the lowest and the human injury is also the lowest, which is very important for clinical evaluation. In order to achieve the purpose of diagnosis, the quality of the image is very important. Now it is often relied on the radiologist to determine whether the image is qualified. However, different radiologists will have slightly different qualification standards, and it is difficult to achieve the image consistency. In recent years, AI has developed rapidly, and many deep learning models have also been applied to medical imaging. If we can use technology to help identify whether chest X-ray images are qualified, we can increase the quality of the images and improve the work efficiency and quality of medical staff. This study will use the object localization model Faster-RCNN and the image classification model ResNet as the basis to identify whether the chest X-ray image is qualified. In order to achieve this goal, this study uses Faster-RCNN to select the region of interest in the image, and then uses ResNet to identify whether the left scapula, right scapula, cervical and costophrenic angle were qualified in chest X-ray images. In the object localization model, this study uses IoU and Y_IoU as the basis for the accuracy of target location positioning, while in the image classification model, the accuracy, sensitivity, and specificity are used to identify the evaluation indicators for the identification of qualified standards for each part.