利用影像後處理增強醫療影像進行深度學習

ASIAIR > College of Creative Design > Department of Media and Design > Theses & dissertations > Item 310904400/113874

Please use this identifier to cite or link to this item: http://asiair.asia.edu.tw/ir/handle/310904400/113874

Title:	利用影像後處理增強醫療影像進行深度學習 Medical Image Enhancement via Post-processing Skill for Deep Learning
Authors:	蔡楨永 TSAI, JEN-YUNG
Contributors:	數位媒體設計學系
Keywords:	超音波影像;加速度;卷積神經網路;磁振造影;頻譜圖 Ultrasound image;Acceleration;Convolutional Neural Network;Magnetic Resonance Imaging;Spectrogram
Date:	2021-01-28
Issue Date:	2022-10-31 06:39:44 (UTC+0)
Publisher:	亞洲大學
Abstract:	醫療影像辨識利用深度學習 (Deep Learning)，已運用於輔助臨床診斷和醫療研究。近年人工智慧 (Artificial Intelligence) 和深度學習領域的發展，在硬體發展或軟體技術都有重大突破。但是醫療影像資料蒐集不易，影像資料不足的問題影響了深度學習的訓練，容易造成資料分佈不均 (Data Imbalance) 或無法進行深度學習訓練。利用影像後處理的優點，增加資料集的影像數量和影像特徵，以減少學習過程中出現的誤差和資料不平衡誤差的問題。本研究假設，透過影像後處理的方式，讓小規模的醫療影像資料集 (Dataset)，進行深度學習的訓練，才有機會得到良好的分類與預測的結果。研究的目的在於實踐少量資料在深度學習中訓練的可行性，使用影像後處理的方法，進行醫療影像深度學習的模型訓練，提高訓練的精準度 (Precision) 和準確度 (Accuracy)。醫療影像蒐集範圍由簡單生物訊號到昂貴的檢測儀器，結合肩痛、疲勞與下背痛等骨骼肌常見的問題。因此本研究的三種醫療影像範圍如下：1.加速度訊號 (Acceleration Signal)、2.超音波影像 (Ultrasound Image)、3.磁振造影 (Magnetic Resonance Imaging, MRI)。這些影像透過需要經過格式轉換，進行分類辨識和找尋病癥。經過影像後處理的透過翻轉、旋轉、尺寸縮放、高斯雜訊、改變明暗對比等，分別在加速度訊號轉換的頻譜圖、肱二頭肌的超音波影像，以及腰椎MRI進行影像後處理。經過深度學習，產生的結果：輪椅加速度之X軸左右移動特徵比Y軸前後移動明顯，準確率達83.3 %。肱二頭肌超音波影像35 Hz與50 Hz的特徵更為明顯，AUC值10分鐘35 Hz為0.834與 50 Hz達到0.814。腰椎MRI在「550-aug」組，預測椎間盤突出的準確率有92.35 %，由此證得影像後處理對於小型資料集的深度學習是非常重要。證實其可行性與本研究的假設是一致的，進行影像後處理可幫助醫療影像進行深度學習。 Medical image recognition of deep learning has been utilized in biomedical research and assistant clinical diagnosis. In recent years, deep learning developments of artificial intelligence have made major progress in hardware technology development and software programming development. The problem of insufficient medical image dataset due to uneven data distribution and data imbalance will affect the deep learning network training and may lead inability to conduct deep learning. The hypothesis of this research is to allow the small-scale medical image datasets to be trained in deep learning through the medical image post-processing to obtain effective image classification and prediction results. The purpose of the research is to practice the feasibility of training a small-scale dataset for deep learning and use image post-processing methods to assistance train deep learning models to improve image recognition precision and accuracy. The scope of medical image collection ranges from simple biological signals to complex nuclear magnetic resonance imaging. It could be combined the common soreness problems of modern people with these medical examinations such as wearable sensors with three-axis accelerometers, ultrasonography machine in clinics, and huge and complex equipment as magnetic resonance imaging in the medical centers. Therefore, this study focuses on the three medical signal and image: 1.Acceleration Signal, 2.Ultrasound Image and 3.Magnetic Resonance Imaging (MRI). These images need to undergo format conversion for classification identification and search correction. Through the image post-processing technique, such as rotation, size scaling, Gaussian noise, changing the contrast of images. The acceleration signal converted to the spectrogram, the ultrasound image of upper biceps, and Lumbar MRI use CNN model to training image recognition through deep learning. The result: the X-axis left and right movement characteristics of the wheelchair acceleration It moves forward and backward more obviously than the Y-axis, with an accuracy rate of 83.3 %. The 35 Hz and 50 Hz characteristics of the biceps ultrasound images in the body are more obvious. The AUC value of 10 minutes at 35 Hz is 0.834 and 50 Hz reaches 0.814. Lumbar MRI in the "550-aug" group, the accuracy of predicting lumbar intervertebral disc herniation is 92.35 %. The medical image uses the skill of post-process to improve the deep learning dataset. These results show better performance of the small-scale dataset confirming the feasibility. It is consistent with the hypothesis of this research.
Appears in Collections:	[Department of Media and Design] Theses & dissertations

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	65	View/Open

Loading...