| 摘要: | 醫學影像技術一直是用來檢測生物細胞組織或是人體組成與內部結構的有效方案，其中包含了核磁共振造影、超音波影像、X光攝影、電腦斷層掃描與顯微成像系統等，並且被廣泛應用到臨床醫學診斷。然而在影像之分析上多是使用人工的方式，因此當需要處理複雜且大量的醫療影像判斷時有其極限與時間消耗上的問題存在。近來由於人工智慧與大數據等技術在影像的迅速發展，已有許多相關醫學上的應用。不僅可以對醫學影像中的微病徵進行初步的判讀，也能進一步描繪出相關病徵的輪廓，可輔助醫生在訊息量龐大且複雜的影像中找出人眼難以察覺及較易於忽略的病徵。
本研究論文中使用遮罩區域卷積神經網路技術進行一系列的醫學影像物件偵測應用，分別建立了針對顯微血球細胞與胸部X光影像的偵測系統與判讀模型，本模型同時繼承了快速區域捲積神經網絡機制達到可針對目標物體辨識與座標標定，同時結合實例分割的方法，完成自動分割和構建圖像中每個目標物件之像素級前景與背景，以利進行可視化邊界框中的物體邊緣，最後成功建立了顯微血球細胞及肺部病徵判定之像素級遮罩模型系統。
本研究也針對血球細胞與胸部X光的肺結節病徵建立了新的資料庫，再經過對模型做微調後，結合遷移學習的技術加速模型的訓練，取代傳統的選擇性搜尋演算法提高了物件辨識率與生成預選框速度。於本研究的結果預測中，預測禎數可高達15每秒，血球細胞與肺結節病徵分類中計算出的平均精度(Mean Average Precision, mAP) 分別為0.931與0.911，而在實例分割則以併交比(Intersection over Union, IoU )計算資料集中生成的血球細胞與肺部X光病徵遮罩的準確率，結果值可達0.961與0.928，足以證實本研究所建立的偵測系統將能夠有機會實際協助醫生於臨床應用上進行快速有效的醫療診斷協助。
;Medical imaging technology, the most effective method for detecting cell tissue or representations of the interior of a body, including radiography, ultrasound, computerized tomography (CT), magnetic resonance imaging (MRI), cardiac angiography and optical microscopy, etc. plays a vital role in clinical analysis.
However, the analysis of images mostly regarded as an operator-dependent method is time-consuming and has narrow limitations when it is necessary to process complex and large amounts of medical images for medical diagnosis.
Recently, the rapid development of technology such as artificial intelligence and big data in imaging has triggered a vigorous growth in medical imaging applications that assist doctors in making initial interpretation of barely imperceptible symptoms and even primary diagnosis for the relevant symptoms from those complicated and large data.
In this paper, Mask Region-based Convolutional Neural Network (Mask RCNN) applied a series of biomedical image object detection for establishing the detection system and the deep learning model for blood cells and chest X-rays respectively.
The model also equipped with a faster convolutional neural network (Faster RCNN) which fulfills the goal of target object identification and coordinate calibration.
Simultaneously, integrated with the method of instance segmentation for automatic segmentation and the construction of pixel-level foreground and background for each target object, the pixel-level mask model system was established successfully to visualize the object counter in the bounding box to make a medical judgment by blood cells and symptoms of the lung.
Moreover, the study organized a new database for blood cells and pulmonary nodules of chest X-ray, then fine-tuned the model and applied with transfer learning to accelerate the training speed, and replaced traditional selective search with Region Proposal Network (RPN) to enhance the recognition rate and the speed of selecting of regions of interest. In this study, the frame per second can be as high as 15 images per second, and the mean average precision calculated in the classification of blood cells is 0.931 and that of pulmonary nodules is 0.911, respectively. While in instance segmentation, intersection over union was applied to calculate the accuracy of the symptom mask of blood cells and chest X-ray which could reach 0.961 and 0.928, respectively. Eventually, the result provided the necessary empirical evidence that the detection system established by the study will realize the vision to assist health professionals with the rapid and effective clinical diagnosis of clinical application. |
