Prenatal brain MRI samples for development of automatic segmentation, target-recognition, and machine-learning algorithms to detect anatomical structures

Introduction

Sample sorting can be useful for clinical research seeking to measure the feasibility of new ideas and to develop new technology. MaZda software (http://www.eletel.p.lodz.pl/programy/mazda/) makers and developers have shown that they are listening to their users by continuing to release updates and new versions^1,2. The samples provided with the manuscript herein were collected and sorted especially for testing upcoming versions of MaZda. The aim is to continue to collaborate with MaZda software engineers in order to code target-recognition semantics and eventually build ideal algorithms for automatic segmentation of the prenatal brain. It is important to also note that it is not an easy task: to deconstruct the scientific knowledge acquired by radiologists after several hours of practice to master the skills of diagnostic imaging. Moreover, we have recently tested MaZda version 4.6 and version 5.0 and made some recommendations to the software engineers^3–5. In reaction to our need, the MaZda team announced an upcoming version called qMaZda, being codeveloped with Weka (www.eletel.p.lodz.pl/pms/SoftwareQmazda.html; www.cs.waikato.ac.nz/ml/weka). We are expecting to see some improvements in qMaZda.

Methods

This dataset was created to improve the efficacy of MaZda. Sample collection was approved by the Research Ethics Committee of the Medical University of Lodz (permit number: RNN/213/13/KE). Subjects were informed with a written statement of consent for research and publication. As per agreement, personal information was removed from the original specimens.

In terms of subject demographics and phenotypes, the background of the patients was consistent with the majority of the Polish population. In 2015, the World Health Organization (WHO) reported 2.68 million neonatal deaths (WHO fact sheet on congenital anomalies, updated September 2016: www.who.int/mediacentre/factsheets/fs370/en/). The estimate of children born with at least one congenital malformation is about 2–3% worldwide (www.who.int/genomics/anomalies/en/Chapter02.pdf). In Poland, the prevalence rate of birth defects was estimated at 52 to 53 per 1000 live births (http://www.marchofdimes.org/materials/global-report-on-birth-defects-the-hidden-toll-of-dying-and-disabled-children-full-report.pdf). Known birth defects can be detected early in pregnancy using non-invasive and/or invasive techniques^6–9. Some can even be treated in utero^10,11. There flows the rationale behind this collation of fetal magnetic resonance imaging (MRI) data to improve the efficacy of MaZda.

The enrolled subjects underwent MRI examination for the purpose of investigating suspected congenital, obstetrical, and placental anomalies that could not be detected by routine ultrasound and genetic amniocentesis. Volunteers who donated fetal MRI samples to create this dataset were in need of fetal, obstetrical or placental care. The criteria for inclusion and exclusion were as follows: 1) 1.5T or 3T MRI; 2) all three anatomical planes were scanned (axial, coronal, and sagittal); 3) individual cross-sectional images are “usable” (i.e., not heavily degraded by noise and artifacts, motion blur, or uncontrollable movement of fetal head); 4) visible fetal brain with no significant malformation; 5) thalamus, gray matter, white matter, and ventricles are also visible (Figure 1). The request to collect MRI scans was sent long after MRI examination was performed. Hence, MRI examination was not prescribed for the purpose of creating this dataset. After looking at 1358 MRI scans in two teleradiology databases, we manually selected 6 patients who had undergone 1.5T MRI examination at Barlicki University Hospital (Łódź, Poland) and 11 patients who had undergone 3T MRI examination at Polish Mother’s Memorial Hospital-Research Institute (Łódź, Poland).

Figure 1. Schematic view of the clinical arrangement of the data.

Fetal magnetic resonance imaging studies were extracted from compact discs and sorted by gestational age of pregnancy and anatomical plane of the mother.

Electro-radiology technicians performed fetal MRIs, as per details provided on the prescription and hospital regulations¹². Hence, we did not have control of MR scanner settings. The technicians stored the MRIs on compact discs (CDs). By default, MaZda version 4.6 and 5 are lacking an automatic samplesort (sorting algorithm) to extract images and arrange them into folders and subfolders. An option was to create a plug-in especially written for MaZda. Due to time consumption, we used other software to carry out image acquisition (Micro Dicom 0.9, Dimensions 2, Sante Dicom 4, Photoshop CS6 64-bit Extended)⁵. For the extraction of Digital Imaging and Communications in Medicine (DICOM) data, we used the 64-bit portable version of Micro Dicom 0.9.1. Most CDs could be accessed with Micro Dicom or Photoshop CS6 64-bit Extended. We used the rescue feature in Sante Dicom 4 to recover the data and export them in DICOM format. Micro Dicom was the preferred samplesort, because it also allowed image selection, batch conversion, and export of DICOM files as 32-bit BMP. Sorting of the sample by MR strength (3T, 1.5T) was carried out with Dimensions 2. Clinical arrangements of the sample by gestational age of pregnancy and anatomical plane of the mother were carried out manually. The primary goal was to select images that had clearly identifiable anatomical regions such as gray matter, white matter, ventricles, and thalamus. It was not possible to complete the task with the MaZda version 5 package, as the available algorithms were lacking automatic segmentation to detect anatomical structures of the fetal brain. Additionally, details about in-depth sorting as well as MRI specifications and file formats are provided in the methods of the cited article⁵.

Ethics and informed consent

Permission to collect samples was approved by the Research Ethics Committee of the Medical University of Lodz. Written informed consent was obtained from all subjects (permit number: RNN/213/13/KE).

Data and software availability

Dataset 1: Fetal MRI data. 1.5/3T samples were manually sorted by gestational age and anatomical plane. Format: 32-bit BMP. doi, 10.5256/f1000research.10723.d150296¹³

MaZda Package v5 RC HG available from: http://dx.doi.org/10.17632/dkxyrzwpzs.1

Micro Dicom 0.9: www.microdicom.com/downloads.html

Dimensions 2: www.skwire.dcmembers.com/wb/pages/software/dimensions-2-folders.php

Sante Dicom 4: www.santesoft.com/downloads.html

Photoshop CS6 64-bit Extended: https://helpx.adobe.com/photoshop/using/dicom-files.html