現今的大腸鏡檢查與影像科學的研究中,主要都是以增進腺瘤偵測率(Adenoma Detection Rate) 為目標,比如將深度學習於物件偵測(Object Detection) 的技術應用在瘜肉偵測上,讓大腸鏡檢查醫師能減少錯失的瘜肉,同時增加腺瘤偵測率。醫生對於切除瘜肉的恰當位置與時機通常是以自身經驗判斷,而每位醫生做瘜肉切除手術所花費的時間也因此有所不同,對於初學大腸鏡檢查的醫生來說,瘜肉切除的位置沒有一定的標準與時機,卻需要一次次在觀摩有經驗者的臨床學習中學習。因此,在本研究中,我們利用單鏡頭大腸鏡的二維影像訓練自監督學習(Self-Supervised Learning) 深度卷積神經網路,提出了大腸深度估測(depth estimation) 的方法,並結合瘜肉偵測的技術開發一套瘜肉定位系統,此外,我們嘗試以數學統計的方式定義出適當的瘜肉切除位置,藉此系統我們希望能增進初學醫師對於瘜肉切除位置的敏感度,減少繁冗的訓練時程。訓練出的瘜肉偵測器在我們的測試數據集中達到92%的mAP@0.5,且深度估測器在我們的測試數據集中達到0.12的絕對相對誤差(Absolute Relative Difference)。
Research on colonoscopy and computer vision in recent mainly focuses on improving Adenoma Detection Rate (ADR). One of the applications of object detection in deep learning is polyp detection that is a computer-aided detection (CADe) tool to reduce the polyp miss rate for colonoscopists and improve ADR. It usually depends on the doctor’s experiences to decide the appropriate position and time while doing polyp removal, so the lasting time every colonoscopist spending on removing a polyp is different. For beginners, there is no absolute standard of removal polyp position, they need to accumulate experience and observing by following experts’ clinical operations again and again. In this thesis, we use monocular colonoscopy 2D images to train a self-supervised learning network, providing a depth estimator estimating depth in colon images. We integrate the technique of polyp detection and then develop a polyp positioning system. In addition, we try to define the adaptive polyp removal position objectively. The system helps beginners enhance the sensitivity of polyp removal positions and reduce the personnel training time. The polyp detector reaches 92% mAP@0.5 in our test dataset and the depth estimator has 0.12 absolute relative difference.