Related Experiment Video
Updated: Nov 5, 2025
10:40
Label-free, High-Resolution 3D Imaging and Machine Learning Analysis of Intestinal Organoids via Low-Coherence Holotomography
Published on: August 12, 2025
906
Unsupervised colonoscopic depth estimation by domain translations with a Lambertian-reflection keeping auxiliary task
Hayato Itoh1, Masahiro Oda2, Yuichi Mori3,4
1Graduate School of Informatics, Nagoya University, Furo-cho, Chikusa-ku, Nagoya, 464-8601, Japan. hitoh@mori.m.is.nagoya-u.ac.jp.
Summary
This study presents an unsupervised depth-estimation method for colonoscopy, crucial for developing computer-aided diagnosis (CAD) systems. The technique accurately extracts 3D colon structures from 2D images, overcoming limitations of existing methods.
Area of Science:
- Medical Imaging
- Computer Vision
- Computational Geometry
Background:
- Three-dimensional (3D) structure extraction from 2D colonoscopic images is vital for computer-aided diagnosis (CAD) systems.
- Existing depth-estimation methods are unsuitable for colonoscopy due to device limitations and lack of ground-truth depth data, hindering supervised learning.
- Accurate 3D information extraction is essential for improving diagnostic capabilities in colonoscopy.
Purpose of the Study:
- To develop a novel, unsupervised, and accurate depth-estimation method for colonoscopic images.
- To overcome the limitations of existing depth-estimation techniques in the context of colonoscopy.
- To enable the development of advanced CAD systems by providing reliable 3D structural information.
Main Methods:
- Proposed an unsupervised depth-estimation method utilizing a Lambertian-reflection model as an auxiliary task for domain translation between real and virtual colonoscopic images.
- Employed the Lambertian-reflection assumption to enhance depth estimation accuracy.
- Conducted qualitative evaluations against state-of-the-art unsupervised methods and quantitative evaluations using a measuring device and a novel 3D reconstruction technique.
Main Results:
- Achieved accurate depth estimation with an average error of less than 1 mm for regions near the colonoscope in quantitative evaluations.
- Demonstrated that the auxiliary task effectively mitigates the impact of specular reflections and colon wall textures on depth estimation.
- Produced smooth, noise-free depth estimations, validating the proposed method's efficacy.
Conclusions:
- Developed an accurate unsupervised depth-estimation method using domain translation with an auxiliary task.
- The method is valuable for analyzing colonoscopic images and advancing CAD system development.
- Successfully extracts precise 3D information from colonoscopic images, enhancing diagnostic potential.
