使用內視鏡影像進行大腸三維重建之研究

Abstract

大腸內視鏡被認為是最有效篩檢大腸癌的方法，可以使存活率大於90%。然而，即便進行了大腸內視鏡的檢查，還是會有腺瘤漏診的情形。因此，本研究提出一個在不需要增加原內視鏡體積的條件下，進行三維腸道重建的方法，希望藉由立體影像的重建，輔助醫生判斷大腸內視鏡檢查時遺漏的區域，減少腺瘤漏診的問題。 本研究先以深度預測條件對抗生成網路 (Conditional Generative Adversarial Network) 產生膠囊內視鏡的RGB圖片對應的深度圖。接著使用RGB圖片、深度圖、內視鏡位置與姿態等三個資訊，利用截斷有號距離函數 (Truncated Signed Distance Function，TSDF) 為主要重建演算法，進行腸道的三維重建。除了原始的TSDF演算法之外，本研究亦結合了兩種優化方式：Distance aware slow saturation (DASS) ，可以針對融合不同視角距離腸壁的遠近而選擇合適的更新資訊；Laplacian smoothing，可以在不增加記憶體用量的條件下，讓重建結果顯得更平滑、更接近真實腸道的樣貌。

Colonoscopy is widely recongnized as the most effective method for screening colorectal cancer (CRC), potential enhancing surcial rates to over 90%. Despite this, there remains a possibility of overlooked adenomas even with colonoscopy examination. In response, this thesis introduces a method for 3D colon reconstruction without increasing the endoscope’s size. By providing 3D reconstructed models, we aim to assist physicians in identifying overlooked areas during colonoscopies, thereby reducing the issue of missed adenomas. In our approach, we first generate depth maps corresponding to capsule endoscope RGB images using a conditional generative adversarial network (CGAN). These RGB images, depth maps, and the enddoscope’s pose and position are then used to perform 3D reconstruction with the truncated signed distance function (TSDF) as the primary algorithm. In addition to the standard TSDF, our approach includes two optimization strategiew: distance aware slow saturation (DASS), which allows for the selection of appropriate updating information based on the relative distance from various viewpoints to the colon wall, and Laplacian smoothing, which smoothens the reconstruction outcome to more closely resemble the actual colon appearance, all without increasing memory use.