Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Internal Organ Localization Using Depth Images

Eytan Kats, Kai Geißler, Jochen G. Hirsch, Stefan Heldman, Mattias P. Heinrich

TL;DR

This paper addresses reducing MRI setup time by automating patient positioning through depth-based localization of internal organs. It introduces a learning-based framework that infers approximate organ locations from simulated depth images derived from approximately 10,000 full-body MRI scans in the NAKO dataset, predicting 11 organs via a U-Net that outputs multi-label masks. Training uses a Dice and binary cross-entropy loss, with depth inputs projected in the coronal plane and evaluated against both manually labeled and TotalSegmentator-generated masks, showing most organs achieve a 95th DOE below 30 mm and favorable Dice/ASSD performance, though bladder and left liver lobe localization remain more challenging. The results demonstrate feasibility and potential integration into MRI workflows, while highlighting domain gap between simulated depth data and real-world RGB-D cameras and suggesting directions for domain adaptation and 3D shape prediction.

Abstract

Automated patient positioning is a crucial step in streamlining MRI workflows and enhancing patient throughput. RGB-D camera-based systems offer a promising approach to automate this process by leveraging depth information to estimate internal organ positions. This paper investigates the feasibility of a learning-based framework to infer approximate internal organ positions from the body surface. Our approach utilizes a large-scale dataset of MRI scans to train a deep learning model capable of accurately predicting organ positions and shapes from depth images alone. We demonstrate the effectiveness of our method in localization of multiple internal organs, including bones and soft tissues. Our findings suggest that RGB-D camera-based systems integrated into MRI workflows have the potential to streamline scanning procedures and improve patient experience by enabling accurate and automated patient positioning.

Internal Organ Localization Using Depth Images

TL;DR

This paper addresses reducing MRI setup time by automating patient positioning through depth-based localization of internal organs. It introduces a learning-based framework that infers approximate organ locations from simulated depth images derived from approximately 10,000 full-body MRI scans in the NAKO dataset, predicting 11 organs via a U-Net that outputs multi-label masks. Training uses a Dice and binary cross-entropy loss, with depth inputs projected in the coronal plane and evaluated against both manually labeled and TotalSegmentator-generated masks, showing most organs achieve a 95th DOE below 30 mm and favorable Dice/ASSD performance, though bladder and left liver lobe localization remain more challenging. The results demonstrate feasibility and potential integration into MRI workflows, while highlighting domain gap between simulated depth data and real-world RGB-D cameras and suggesting directions for domain adaptation and 3D shape prediction.

Abstract

Automated patient positioning is a crucial step in streamlining MRI workflows and enhancing patient throughput. RGB-D camera-based systems offer a promising approach to automate this process by leveraging depth information to estimate internal organ positions. This paper investigates the feasibility of a learning-based framework to infer approximate internal organ positions from the body surface. Our approach utilizes a large-scale dataset of MRI scans to train a deep learning model capable of accurately predicting organ positions and shapes from depth images alone. We demonstrate the effectiveness of our method in localization of multiple internal organs, including bones and soft tissues. Our findings suggest that RGB-D camera-based systems integrated into MRI workflows have the potential to streamline scanning procedures and improve patient experience by enabling accurate and automated patient positioning.

Paper Structure

This paper contains 4 sections, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. Materials and methods
  3. Results
  4. Discussion

Figures (2)

  • Figure 1.1: Detection Offset Error (DOE) for anatomical structures in 30 manually segmented MRI scans. The value above each box corresponds to 95 percentile of DOE.
  • Figure 1.2: Comparison of predicted and ground truth segmentations for internal organs. To enhance visualization clarity and address potential overlaps of 2D projections, organs are presented in a separate images. From left to right: depth image derived from MRI scan, liver, vertebrae and urinary bladder, lungs, heart and kidneys, spleen and hips, femurs and pancreas. Red - ground truth mask, green - model prediction, yellow - overlap between ground truth and predicted mask.