深度視覺運動控制之磁控膠囊內視鏡自動牽引

Abstract

近年來，由於大腸相關的病灶在國內都是排列第一的，而根據研究報告早期大腸鏡的篩檢有助於提高與結腸癌相關之存活率。在傳統的內視鏡檢查會因其固有的操作複雜性而產生了門檻的限制，而我們實驗室所使用的磁控膠囊內視鏡則相較於傳統內視鏡有了許多優點，例如:舒適性、低壓迫性、遠程控制等優點。 近年來各行各業陸陸續續導入AI技術，在醫學領域中也不例外，為了減少內視鏡醫師的工作負擔和操作時間，我們實驗室的學長已有開發透過腸鏡鏡頭獲取之影像處理、訓練深度學習模型、梯度計算相關應用達成由直腸至盲腸之全自動牽引之可行性研究。在本篇論文中，我期望使用另一種的深度學習模型將模型做改良並且使模型複雜化降低。我使用了一種強化學習(Reinforcement Learning，RL)之變型深度視覺運動控制(Deep Visuomotor Control，DVC)來做改良並在模擬的環境裡達到更好的任務效果。 強化學習雖然在機器學習上能完成許多任務，但在現實環境中機器人並不像人類有著聽覺、觸覺、味覺和嗅覺，而深度視覺運動控制則是開發了一種方法基於強化學習將原始圖像透過Convolutional Neural Networks(CNN)將強化學習的前處理變成監督學習，而處理後的輸入則作為訓練強化學習Policy使一系列需要在實現了在現實環境視覺和控制之間的協調。在本篇論文期望使用這種方法實現內視鏡在腸道自動牽引的高維連續動作。

Recently, colorectal diseases have emerged as the most prevalent health concern domestically. Early colonoscopy screenings, as research suggests, enhance survival rates for colon cancer patients. Traditional endoscopy, due to its inherent complexity, poses significant challenges. Our lab employs magnetically-controlled capsule endoscopy, providing advantages like greater comfort, reduced invasiveness, and remote control capabilities. Various sectors are integrating AI technology, with medicine being no exception. To lessen the burden on endoscopists, previous work in our lab has developed a deep learning model for automating rectal-to-cecal traction, leveraging image processing and gradient computations. This paper seeks to improve and simplify this model using a variant of Reinforcement Learning (RL), termed Deep Visuomotor Control (DVC), achieving superior performance in a simulated environment. RL is proficient at various machine learning tasks, but real-world robots lack human senses. DVC addresses this by transforming RL preprocessing into supervised learning via Convolutional Neural Networks (CNN). This transformed input is used as an RL training policy, enabling coordinated actions in realistic vision and control scenarios. This paper aims to utilize this approach for implementing automatic endoscope traction in the intestines.