本文選題：腰椎　切入點：卷積神經網絡　出處：《南京理工大學》2017年碩士論文　論文類型：學位論文

【摘要】：隨著數(shù)字圖像技術與計算機視覺技術的不斷發(fā)展,使用計算機技術手段對醫(yī)學圖像數(shù)據(jù)進行處理和分析來輔助醫(yī)生診斷疾病已逐漸得到普及。本文針對腰椎不穩(wěn)癥這一普遍而嚴重的健康問題,提出了一種腰椎不穩(wěn)癥輔助診斷方法:基于學習的腰椎檢測和跟蹤方法。該方法使用腰椎的數(shù)字視頻影像(DVF)進行處理與分析,主要研究工作如下:(1)針對目前腰椎跟蹤方法中初始狀態(tài)的腰椎目標均需通過人工標注的問題,提出了一種基于卷積神經網絡的腰椎檢測方法。首先對原始DVF圖像進行對比度拉伸和去噪預處理,增加DVF圖像的清晰度;在離線訓練階段,使用大量腰椎樣本圖像來訓練卷積神經網絡分類器;檢測時,利用霍夫變換在二值化的DVF圖像中尋找到腰椎的邊角點以得到腰椎的角度參數(shù),再使用腰椎解剖統(tǒng)計數(shù)據(jù)獲得初始候選檢測區(qū)域的邊界框;之后按邊界框從經預處理的DVF圖像中提取初始候選檢測區(qū)域送入卷積神經網絡分類器當中獲得檢測結果。該方法在實驗中表現(xiàn)出了十分高的腰椎檢測準確率,能夠實現(xiàn)對精度較嚴格的場景下的應用。(2)針對當前許多跟蹤算法在對腰椎進行跟蹤時的魯棒性較差的問題,提出了一種可在線更新的基于棧式自動編碼機的腰椎跟蹤方法。在離線訓練時,獲得能夠表述通用物體的深層特征;在線跟蹤時,以粒子濾波為框架,使用自動編碼機獲得當前幀的跟蹤結果;最后再對神經網絡權值參數(shù)進行在線更新。該方法有效提高了算法的魯棒性并減少了跟蹤漂移的可能,在實驗中展現(xiàn)出了很強的腰椎識別能力。以上技術能夠在DVF影像對比度較低且較為模糊的情況下將部分腰椎檢測出來,并展現(xiàn)出了很強的識別能力,在跟蹤過程中可準確發(fā)現(xiàn)腰椎不穩(wěn)癥狀,可作為腰椎不穩(wěn)癥臨床診斷中非常有效的輔助手段。
[Abstract]:With the development of digital image technology and computer vision technology, The use of computer technology to process and analyze medical image data to assist doctors in diagnosing diseases has become increasingly popular. In this paper, an auxiliary diagnosis method of lumbar vertebrae instability is proposed, which is based on learning and tracking of lumbar vertebrae, which is processed and analyzed by digital video image of lumbar vertebrae (DVF). The main research work is as follows: (1) aiming at the problem that the initial status of lumbar vertebrae in the current lumbar tracking method needs manual labeling, A novel lumbar spine detection method based on convolution neural network is proposed. Firstly, the contrast stretching and denoising preprocessing of the original DVF image is carried out to increase the clarity of the DVF image. The convolutional neural network classifier is trained with a large number of lumbar vertebrae samples, and the angle parameters of the lumbar vertebrae are obtained by using Hoff transform in the binary DVF image to find the side corner of the lumbar vertebrae. The boundary frame of the initial candidate detection area was obtained by using lumbar anatomical statistical data. Then, according to the boundary frame, the initial candidate detection area is extracted from the pre-processed DVF image and sent to the convolutional neural network classifier to obtain the detection results. To solve the problem that many current tracking algorithms have poor robustness in tracking the lumbar vertebrae. In this paper, an on-line updating method of lumbar spine tracking based on stack automatic coding machine is proposed. When training offline, the deep features of common objects can be expressed, and the particle filter is used as the framework for on-line tracking. The tracking result of the current frame is obtained by using the automatic coding machine. Finally, the weights of the neural network are updated online. This method can effectively improve the robustness of the algorithm and reduce the possibility of tracking drift. These techniques can detect parts of lumbar vertebrae in the case of low contrast and blur of DVF image, and show strong recognition ability. The unstable symptoms of lumbar vertebrae can be found accurately in the process of tracking, and can be used as a very effective auxiliary method in the clinical diagnosis of lumbar instability.
【學位授予單位】：南京理工大學
【學位級別】：碩士
【學位授予年份】：2017
【分類號】：TP391.41;R445;R681.5

【參考文獻】

相關期刊論文 前1條

1 Wang Guohong;Tan Shuncheng;Guan Chengbin;Wang Na;Liu Zhaolei;;Multiple model particle flter track-before-detect for range ambiguous radar[J];Chinese Journal of Aeronautics;2013年06期

，

本文編號：1587389