Title: Deep-learning based image exposure correction and depth estimation for 3D cartography in endoscopy
Abstract:
This thesis investigates the application of deep learning techniques to improve 3D reconstruction in endoscopy. A passive vision framework is proposed to enhance monocular video data by means of photometric correction and self-supervised depth estimation, addressing challenges related to uneven illumination and limited geometric information. The contributions include the development of Endo4IE, a synthetic dataset specifically designed to support the training of exposure correction models, as well as the implementation of deep-learning methods capable of recovering structural details in poorly illuminated regions. In addition, an illumination-invariant and self-supervised depth estimation model is introduced to infer scene geometry from monocular sequences without requiring ground-truths. These modules have been integrated into a 3D reconstruction pipeline enabling the generation of dense surfaces with coherent shapes. Extensive experiments on both synthetic and real colonoscopic data demonstrate significant improvements in i) the image contrasts, ii) the accuracy of the 3D camera trajectory tracking, and iii) the quality of the surface construction. These results all contribute to a more reliable and geometry-aware visualization in clinical endoscopy.
Jury Composition:
Rapporteurs:
David Fofi, PU, Université de Bourgogne, Laboratoire ImViA
Jean-Bernard Hayet, CIMAT, Guanajuato, Mexico
Examiners:
Mariel Alfaro-Ponce, Tecnológico de Monterrey, Campus Mexico City
Marie-Odile Berger, Directrice de Recherche INRIA, LORIA
Thesis Co-directors:
Christian Daul, PU, Université de Lorraine, CRAN
Gilberto Ochoa-Ruiz, Tecnológico de Monterrey, Campus Guadalajara
Teams link: click here