Deep Learning based Model Observers for Multi - Modal Imaging

Naderi Boldaji H, Patwari M, Reymann M, Gutjahr R, Raupach R, Maier A (2021)

Publication Language: English

Publication Type: Conference contribution, Conference Contribution

Publication year: 2021

Event location: Leuven, Belgium BE

Abstract

Model Observers (MO) are useful tools for assessing task-based image quality in medical imaging systems. Detection and characterization of anatomical structures in clinical imaging are some of the most significant tasks of MOs. Various linear model observers have been widely applied for these tasks with great success. However, we are unable to deploy such observers without a high prior knowledge of the task, the signal, and the background. Additionally, most model observers currently used in medical imaging research are used to optimize single modality systems such as standalone CT or PET scanners. In this paper, we aim to solve these problem by means of a Convolutional Neural Network (CNN). We employ supervised learning with CNNs to approximate the Human Observer (HO) for a signal-known-exactly and background-known-statistically (SKE/BKS) signal detection task. We use this observer in a novel multi - modality imaging setting. First, we propose a CNN-based anthropomorphic model observer to predict human observer detection performance for PET images. Then, we suggest another CNN-based MO for multi - modal PET and CT images. We trained both networks using $160$ sets of joint PET/CT images of a dataset from the MICCAI2020 HECKTOR challenge, and tested on $41$ patients from the same dataset. Our CNN - MOs have high accuracy in detecting the lesions present in the image (AUC = 0.95). Furthermore, our MOs degrade monotonically with increase in noise levels, similar to how a HO would. Including both modalities into an MO results in improved lesion detection performance, especially in the presence of image noise, compared to single modality CNN - MOs.

Authors with CRIS profile

Mayank Patwari Lehrstuhl für Informatik 5 (Mustererkennung) Maximilian Reymann Lehrstuhl für Informatik 5 (Mustererkennung) Andreas Maier Lehrstuhl für Informatik 5 (Mustererkennung)

Involved external institutions

Siemens Healthineers DE Germany (DE)

How to cite

APA:

Naderi Boldaji, H., Patwari, M., Reymann, M., Gutjahr, R., Raupach, R., & Maier, A. (2021). Deep Learning based Model Observers for Multi - Modal Imaging. In Proceedings of the 16th International Meeting on Fully Three-Dimensional Image Reconstruction in Radiology and Nuclear Medicine. Leuven, Belgium, BE.

MLA:

Naderi Boldaji, Hamidreza, et al. "Deep Learning based Model Observers for Multi - Modal Imaging." Proceedings of the 16th International Meeting on Fully Three-Dimensional Image Reconstruction in Radiology and Nuclear Medicine, Leuven, Belgium 2021.

BibTeX: Download