基於深度學習在內視鏡檢查中估計覆蓋率之研究

Abstract

近年來，隨著運算資源的不斷進步，圖形處理器(Graphics Process Unit，GPU)、張量處理單元(Tensor Process Unit，TPU)等專為深度神經網路所研發的AI運算引擎出現，加上物聯網裝置爆炸性的增加，大數據的取得已不是件難事，也因為這種種原因，使得各行各業陸陸續續導入AI技術，當然在醫學影像領域中也不例外，例如: X光攝影、超音波影像、電腦斷層掃描(Computed Tomography，CT)、核磁共振造影（Magnetic Resonance Imaging，MRI）的病徵辨識。 我們知道在內視鏡檢查品質指標中，有包含:盲腸到達率(Cecal Intubation Rate，CIR)、腺瘤偵測率(Adenoma detection rate，ADR)等等……，如果內視鏡檢查過程中可以獲取其深度資訊，不但可以幫助內視鏡醫生提升上述所提之內視鏡檢查品質指標，還可以透過RGB-D圖片將腸道做三維重建，進一步使醫生得知檢查過程中的覆蓋率是否足夠。 在本篇論文中，我將使用多種不同的深度學習網路並應用在大腸內視鏡之深度估計任務當中，來取得腸道之RGB-D 圖像，並比較各模型之優缺點以及做定量評估。此外我還將高質量的深度資訊透過CNN1D模型，訓練出能推測檢查過程之覆蓋率的深度模型，最後使用Truncated Signed Distance Function (TSDF)演算法，實現單幀腸道圖片的三維影像重建。

With the continuous improvement of computing resources for the past few years, such as the Graphics Process Unit and the Tensor Process Unit specially designed for deep learning. Additionally, with the explosive growth of IoT devices, it is no longer difficult to obtain Big data. For these reasons, all walks of life have gradually lead-in artificial technology, the field of medical imaging is no exception. For instance, symptom identification of medical radiography, ultrasonic imaging, Computed Tomography Imaging, and Magnetic Resonance Imaging. We know that the quality indicators of endoscopy include Cecal Intubation Rate, Adenoma Detection Rate, etc. if endoscopists can obtain the depth information during the endoscopy process, it can not only help endoscopists to improve the quality indicators of endoscopy mentioned above but also make 3D modeling of the colon by using the depth information, further inform endoscopists that the coverage during examination is efficient. In this paper, I will use a variety of deep learning models and apply them to the depth estimation task in colorectal endoscopy to obtain depth information. In addition, make a comparison between each model for analyzing benefits, drawbacks and make a quantitative evaluation. Otherwise, I use the high-quality depth information as the CNN-1D model’s training data to train a model which can infer the coverage of the endoscopy. Finally, I use the Truncated Signed Distance Function (TSDF) algorithm to achieve 3D image reconstruction of a single-frame intestinal image.