| MS Seminar

Name of the Speaker: Mr. Rahul G S (EE20S136)
Guide: Prof. Mohanasankar S
Online meeting link: https://meet.google.com/wot-mmyu-ezi
Date/Time: 14th June 2024 (Friday), 4:45 PM
Title: Development of Stereo Endoscopy System for Polyp Measurement and 3-D Reconstruction.

Abstract :

Estimating depth map in endoscopic scenes is crucial for improving surgical precision, navigation, and the visualization of internal structures. It also aids in the 3-D reconstruction of anatomical structures and measuring abnormal tissue sizes. However, traditional endoscopes with a single image sensor cannot provide the necessary depth information. This work discusses the design and development of a stereo endoscopy system that uses a dual camera setup for accurate depth estimation. It discusses the design challenges of creating a stereo endoscopy system and outlines effective strategies to address them. A prototype was developed, including an endoscope, a video processor that connects to a display monitor, and a Virtual Reality-enabled Head Mounted Display Device for enhanced 3-D visualization.

Additionally, this research proposes a new algorithm using a pre-trained Visual Geometry Group Network (VGGNet) model for estimating disparities and accurately measuring the size of abnormal tissues like polyps or lesions. Our algorithm outperforms the widely used Semi Global Matching (SGM) algorithm, achieving an error rate of less than 5% when measuring polyps located 30 mm to 50 mm in depth from the camera.

Furthermore, the talk will discuss the Dual Attention Concatenation Volume Net (DACVNet) for stereo disparity estimation, an improvement over the Attention Concatenation Volume Net (ACVNet) by incorporating a self-attention mechanism. This mechanism allows the model to independently focus on different image sections and recognize patterns across distant pixels. DACVNet demonstrated superior performance over ACVNet and other models by reducing the End Point Error by 7.08%. Qualitative validation using the gastric phantom dataset confirmed the effectiveness of DACVNet in 3-D scene reconstruction, suggesting its potential use in stereo endoscopic applications.