動態影像分析於醫學領域之應用

Abstract

近年來由於硬體技術的進步，以及影像資料的大量累積，電腦視覺領域蓬勃發展，如今已廣泛應用於各領域中，包含工廠自動化、無人駕駛與辨識系統等。隨著人們對醫療品質的要求逐漸上升，電腦視覺也拓展到醫學領域上，透過特殊的醫療攝影設備，結合智慧演算法輔助醫療人員進行診斷，例如藉由斷層掃描進行主動脈剝離診斷，或利用乳房攝影診斷乳房纖維腫瘤等。相較於醫療人員，電腦視覺具有更快、更穩定與更準確的判斷。然而如今應用於醫學領域之電腦視覺技術，主要著重於靜態影像分析，鮮少動態影像分析的例子。因此本研究以開發醫療型眼動儀與穩定跌倒偵測系統為目標，希望藉此將電腦視覺動態影像分析技術應用於醫學領域上。 本研究開發之醫療型眼動儀具備穩定追蹤眼動、準確估測視線與即時分析眼動情形等功能，有別於傳統醫療型眼動儀，無需任何穿戴裝置，且在極短時間內即可量測出受試者視線追蹤能力，以及眼球震顫的嚴重度。系統內部的眼動追蹤演算法，結合基於像素值與像素值之梯度二種特徵進行判定，達成更穩定的追蹤。視線估測法也結合了眼球模型與迴歸分析，降低計算複雜度的同時維持估測精準度。 另一方面本研究設計之跌倒偵測演算法，串聯各種特徵判定方式，包含跌倒時之整體動量、軸心方向與位置、輪廓形狀變化與昏迷可能性等四種方法，計算出六個特徵值，並藉由大量影像資料集測試，調整演算法參數，完成高偵測率且低錯誤警告率的偵測系統，且對運算成本要求極低。

Over the last few years, due to the evolution of hardware technology and the accumulation of image data set, the field of computer vision has developed vigorously. Nowadays, it has been widely used in various fields, such as factory automation, self-driving systems and recognition systems. With the increasing requirements for medical quality, computer vision has also been expanded to the field of medical science, assisting medical personnel in diagnosis through combining special medical photography equipment with intelligent algorithms, such as automatic aortic dissection diagnosis by tomography, or automatic fibroadenoma diagnosis by mammography. Compared with medical personnel, computer vision provides faster, more stable and more accurate diagnosis. However, the computer vision technology applied in the medical field has mainly focused on image analysis. There are only few cases of video analysis in this field. Therefore, this research aims to develop a medical eye tracker and a stable fall detection system, hoping to apply video analysis technology to the medical field. The eye tracker we develop for medical use has the functions of stable eye movements tracking, accurate estimation of gaze direction, and real-time analysis of eye movements. Different from traditional medical eye trackers, it does not require any wearable devices, and is able to measure the gaze tracking ability and the severity of nystagmus of the subjects within very short time. The eye tracking algorithm used in the eye tracker combines intensity-based method with gradient-based method to achieve more stable tracking. In addition, the gaze estimation method combines the three-dimensional eyeball model method with regression analysis, reducing computational complexity and maintaining the accuracy of the estimation meanwhile. On the other hand, the fall detection system designed in our research adopts four feature determination methods to calculate six eigenvalues, including the total momentum, the main axis direction and position, the shape change and the possibility of coma while falling. By testing a large number of image data sets to the algorithm, we optimize the parameters of our detection system to achieve high detection rate, low false alarm rate and extremely low computing cost.