Reconstrucción 3D a escala real a partir de imágenes monoculares de endoscopio.

Martínez Batlle, Víctor
Tardós Solano, Juan Domingo (dir.)

Universidad de Zaragoza, EINA, 2021
Informática e Ingeniería de Sistemas department, Ingeniería de Sistemas y Automática area

Máster Universitario en Robótica, Gráficos y Visión por Computador

Abstract: In today's scientific community, there is a growing research interest in extending augmented reality and autonomous navigation into the human body. The European EndoMapper project aims to solve the problem of real-time 3D mapping of the human colon that will assist colonoscopy, tumour biopsy and other medical procedures. Due to space limitations inside the human bowel, medical endoscopy can only use monocular vision. However, conventional mapping systems are not able to recover the real scale of the environment from monocular images. This intrinsic limitation leads to unknown scale maps, making it difficult to diagnose some diseases of the human colon. In addition, these maps are often distorted, as they suffer from scale drift, which is a common problem in monocular systems.This MEng thesis proposes a solution to monocular real-scale 3D reconstruction inside the human colon. Our approach exploits the controlled lighting inside the human body, where the only light source moves jointly with the camera. This allows us to consider a pseudo-stereo pair formed by the endoscope's light and camera, and use it to achieve real-scale perception on a monocular endoscope. First, we define a model of the illumination and camera of an endoscope, which allows us to understand the imaging process during colonoscopy. Then, this model is adapted and calibrated for a real medical endoscope, using only a sequence of images recorded by the endoscope itself. Finally, we propose a method capable of estimating a dense depth map from a single monocular image, based on the calibration we performed. To evaluate the accuracy of our depth estimation, we conducted experiments with both synthetic and real images of the human colon. With respect to synthetic data, we obtain results with a 7% error, which is less than 3 mm on average. Lastly, we demonstrate that our method can also work with real images, where we estimate dense depth maps that preserve the structure and discontinuities of the human colon.

Tipo de Trabajo Académico: Trabajo Fin de Master