Benchmarking Deep Learning based CT Image Reconstruction Methods

Date: June 15–24, 2020
The sprint will be a virtual event (due to the current pandemic). In addition, we will organize a physical follow-up meeting in Bremen when the situation will allow it.

Organization: Maureen van Eijnatten, Poulami Ganguly, Johannes Leuschner, Maximilian Schmidt

Download as PDF.

General introduction

Deep learning has opened up a wealth of promising methodologies for various image analysis tasks such as image classification, object detection, segmentation, and registration. Moreover, deep learning is being increasingly used to solve inverse problems in computed tomography (CT). In this field, deep learning holds the promise of performing fast and accurate reconstruction of CT images in e.g. high-resolution, limited-angle, or low-dose setups. However, there are still multiple challenges to overcome. Aside from the computational and data-related challenges, solving inverse problems using deep learning possibly requires completely new approaches that allow for the extraction of physically meaningful parameters. In recent years, many different deep learning based methods have been proposed for CT image reconstruction, such as unrolled iterative schemes, learned regularizers, and post-processing of reconstructed images. However, it remains difficult to compare the performance of these methods across different imaging domains and applications due to a lack of generic benchmarking methods and sufficiently large datasets.

During this Code Sprint, we invite a small group of researchers and software developers to exchange knowledge, write code, form new collaborations and connect with industrial partners in the area of (learned) CT image reconstruction. We will focus on two applications with high relevance and impact: 1) reconstruction of low-dose medical CT images, and 2) high-throughput CT image reconstruction and abnormality detection.

Reconstruction of low-dose medical CT images

In the clinical application of CT, high radiation doses are potentially harmful to the subjects. Therefore, better reconstruction methods for low-dose CT acquisitions are highly desirable.

In order to compare (learned) reconstruction techniques for this application, we will use the low-dose parallel beam (LoDoPaB) CT dataset. This dataset contains more than 40 000 two-dimensional CT images and corresponding simulated low-intensity measurements. The ground truth images of this dataset are human chest CT reconstructions taken from the LIDC/IDRI database. Poisson noise is applied to the projection data after simulation to model the low intensity setup.

High-throughput CT image reconstruction and abnormality detection

When using X-ray tomography in high-throughput settings (e.g. scanning multiple objects per second) such as quality control, inspection of products on conveyor belts, or luggage scanning, very few X-ray projections can be acquired of each object. In such settings, it is essential to incorporate a priori information about the object being scanned during image reconstruction. The challenge is to identify and explore how machine learning can be used in the image reconstruction and subsequent decision-making process (e.g. for detecting an internal abnormality in the scanned object) with high computational efficiency.

For this application, annotated CT images of real apples with internal defects will be made available by the innovative fruit sorting company Greefa.

Goal

During the Code Sprint, participants are invited to adapt or implement learned methods to reconstruct CT images based on a medical and/or an industrial training CT dataset (including ground truth reconstructions) that will be provided beforehand. The implemented learned reconstruction methods will be included in the python library DIVαℓ to make them easily accessible to the community. The aim of the Code Sprint is to compare the performance of the implemented methods on a medical and/or an industrial testing CT dataset in a challenge setup.

Output

We aim to publish a comprehensive comparison of the results afterwards. In addition, this Code Sprint will enable easier baseline comparisons when new learned methods emerge. A fair benchmark of learned and classical approaches is an essential next step towards deploying deep learning in medical and industrial CT reconstruction tasks.

Questions for discussion

An important open problem is to define quality measures for reconstructions that consider the needs of the application. For example in medical imaging, the standard metrics (PSNR, SSIM) most likely do not indicate very well whether a doctor could make a better diagnosis based on the reconstruction.