The JNU-IFM dataset for segmenting pubic symphysis-fetal head.

The use of transperineal ultrasound techniques for the assessment of fetal head descent and progression is an adjunct to clinical examination. Automatic identification of parameters based on ultrasound images will greatly reduce the subjectivity and non-repeatability of the clinician's judgment. However, the lack of a pubic symphysis-fetal head dataset hinders the development of algorithms. Here, we present an intrapartum transperineal ultrasound dataset of the Intelligent Fetal Monitoring Lab of Jinan University (named the JNU-IFM dataset), in which intrapartum transperineal ultrasound videos of 78 were recorded from 51 patients. These data were obtained with the Youkey D8 wireless 2D ultrasound probe with its corresponding supporting software by Wuhan Youkey Bio-Medical Electronics Co., Ltd., Wuhan, China. In these videos, 6224 high-quality images with four categories were selected to form the JNU- IFM dataset. These images were labelled using the Pair software and then validated by two experienced radiologists. We hope that this data set can be used in the segmentation of the pubic symphysis-fetal head.

Specifications Table

Value of the Data

The use of intrapartum transperineal ultrasound (ITU) techniques for the assessment of fetal head descent and progression is an adjunct to clinical examination. Compared with subjective judgment and uncertain reproducibility of clinical examination, ITU has the advantages of being well reproducible and objective.

Manual segmentation of symphysis pubis (SP)-fetal head from ITU images for clinical radiologists is considered as the most reliable but extremely time-consuming procedure prone to subjectivity and large inter-observer variability. With the rapid development of artificial intelligence in medical images, automatic measurement algorithms based on ITU images are expected to solve the above problems.

This dataset with ITU images and their labels is useful for developing and evaluating automated SP-fetal head segmentation algorithms and image classification algorithms. Although the SP-fetal head identification plays an important role in computing angle of progression (AoP) to assess the descent of the fetal head during labor, this dataset can also be used as objective and quantitative indicators for evaluating other ITU parameters and tracking their efficacy.

Data Description

This dataset is publicly available at https://doi.org/10.6084/m9.figshare.14371652, which can be downloaded as a zip file. In the unzip file, 78 files are named as “AAAABBCCTDDEEFF” which is the time when the data was obtained. In detail, AAAA, BB, CC, DD, EE and FF represents year, month, day, hour, minute and second, respectively. In the “AAAABBCCTDDEEFF” file, three folders named as “image”, “mask”, and “frame_label.csv” are listed. The “image” folder contains the original images which are saved in the PNG format and named as “AAAABBCCTDDEEFF_G.png” (G indicates which frame the image is in the video). And the “mask” folder contains the labels of the corresponding images in the ``image'' folder and these labels are named as “AAAABBCCTDDEEFF_G_mask.png”. Consequently, the frame number (“G”) and the frame label (3: None, 4: OnlySP, 5: OnlyHead or 6: SPHead) are, respectively, stored in the “frame_id” and “frame_label” columns of the file “frame_label.csv”. It is worth noting that the image in the ``mask'' folder may appear to be an all-black image due to the low label value (SP: pixel value of 7, Head: pixel value of 8, and the remaining pixel values of 0). The numbers of four types of images from each patient are listed in Table 1. There are 6224 images, including 1022 images with the None label, 323 images with the OnlySP label, 1136 images with the OnlyHead label and 3743 images with the SPHead label.

Table 1

Summary of the JNU- IFM dataset. The names of the folder of images obtained from each patient and the corresponding numbers of images in each category are given.

Patient	None	OnlyHead	OnlySP	SPHead	Files
1	11	0	96	0	20190830T11551520190830T11560220190830T115644
2	0	0	9	0	20190904T101559
3	51	0	67	8	20190906T10514520190906T105237
4	17	2	13	41	20190909T155747
5	26	10	3	106	20190909T16145320190909T161601
6	42	0	40	26	20190911T10443720190911T105058
7	0	0	53	0	20190911T111121
8	42	11	26	31	20190916T104520
9	28	0	23	98	20190916T10552620190916T105641
10	11	0	0	63	20190916T110257
11	0	0	0	30	20190916T112312
12	12	0	4	58	20190918T115054
13	11	0	108	75	20190918T12001120190918T12062820190918T120708
14	29	0	26	137	20190918T12334220190918T123437
15	123	0	92	0	20190922T101601
16	0	0	69	0	20190923T173644
17	4	0	68	0	20190923T175155
18	16	10	0	50	20190930T110010
19	4	2	0	66	20191003T173034
20	2	0	1	70	20191008T112326
21	0	0	0	127	20191008T11415920191008T114249
22	11	0	2	62	20191026T195815
23	4	0	0	75	20191108T114950
24	2	0	0	145	20191115T10562320191115T105730
25	2	0	65	7	20191115T110256
26	3	7	0	65	20191115T112747
27	8	0	64	1	20191115T114514
28	11	14	0	121	20191127T11032320191127T110427
29	5	69	0	0	20191127T111516
30	0	2	0	119	20191127T11270020191127T112757
31	14	1	0	132	20191127T11365820191127T113821
32	18	17	9	106	20191129T10313320191129T103822
33	0	5	0	70	20191129T105514
34	17	8	11	118	20191129T11055820191129T110732
35	26	8	14	24	20191203T105250
36	13	0	0	131	20191203T11162720191203T111732
37	16	0	95	39	20191205T10374920191205T103904
38	16	4	0	134	20191208T16454920191208T165241
39	16	19	18	24	20191208T170945
40	26	0	0	49	20191212T102143
41	23	0	7	118	20191212T10320520191212T103310
42	26	0	39	11	20191214T100241
43	25	0	3	46	20191214T103803
44	7	22	47	59	20191218T10474520191218T104848
45	15	44	3	160	20191218T10573520191218T10590920191218T110113
46	66	0	6	0	20191220T102712
47	74	0	1	0	20191220T104055
48	43	0	46	33	20191220T11200220191220T112127
49	30	2	5	115	20191220T11312620191220T113230
50	58	34	3	696	20200103T10262320200103T102728
51	18	32	0	97	20200103T10491920200103T105033

Schematic display generated by angle of progression (AoP) are shown in Fig. 1.

Fig. 1

Transperineal ultrasound to measure the angle of progression (AoP) formed between a straight line drawn along the longitudinal axis of the symphysis pubis (SP) and a line running from the inferior edge of SP to the leading edge of the fetal head. (A)Schematic diagram of calculating AoP; (B) An image with symphysis pubis and fetal head; (C) The segmentation result of the symphysis pubis (red) and fetal head (green); (D) Calculate AoP by elliptic function fitting.

Examples (None, OnlySP, OnlyHead and SPHead) of the original (“Image”) and label (“Mask”) images are shown in Fig. 2. The pixel values corresponding to SP and Head are enhanced (“Enhanced mask”) here to be close to the style marked with Pair software, to facilitate the reader's understanding of the dataset.

Fig. 2

Examples (None, OnlySP, OnlyHead and SPHead) of the original (“Image”), label (“Mask”) and enhanced label (“Enhanced mask”) images.

Experimental Design, Materials and Methods

A total of 78 videos from 51 pregnant women were collected from NanFang Hospital of Southern Medical University between 2019 and early 2020. This study was approved by the Medical Ethics Committee of NanFang Hospital of Southern Medical University (NFCE-2019–024). All authors confirm that we have complied with all relevant ethical regulations.

Transperineal ultrasound examinations were performed in standard B-mode ultrasound using the Youkey D8 wireless 2D ultrasound probe with its corresponding supporting software (Wuhan Youkey Bio-Medical Electronics Co., Ltd., Wuhan, China), which has a 3.53 0.0525 MHz convex probe installed. The spatial resolution of the ultrasound system is specified by the manufacturer to less than 2 mm. The overall geometric inaccuracy of a very similar setup due to inherent technical limitations was measured to be <2.0 mm laterally, <2.0 mm vertically, <2.0 mm longitudinally, and <8.0 mm radially (‘vector length’ or Euclidean ‘3D-distance’; the square root of the sum of squares of the three axes) consisting of random errors (per single measurement point) and systematic errors (effectively, per fraction). The temporal resolution of the device is specified to about 27 Hz.

In order to obtain high-quality images, the transducer was prepped by covering it with a surgical latex glove filled with coupling gel, then the prepped transducer, after applying gel, was placed between labia below the pubic symphysis to obtain a sagittal plane, small adjustments in the form of lateral movements of the probe were made until an image obtained showed clear maternal pelvic (pubic symphysis) and fetal (fetal skull) landmarks that did not show any shadows from the pubic rami [1], [2], [3], [4], [5], [6], [7], [8], [9]. Videos were in the MP4 format, with a resolution of 1920 × 1080. 6224 images are extracted from videos in 10 frames each. The mp4 is lossy compression or noise present in the mp4. However, we cannot remove noise because we export the data using the software that comes with the ultrasound instrument. It is worth noting that the image in the folder ``image'' has been preprocessed basically, and the original interface toolbar and text information of the image have been removed through cutting and overwriting operations. After processing, the resolution is 1295 × 1026. In order to prevent information loss, no downsampling is conducted. At the same time, it has been converted into a grayscale image, which can be read directly.

Ground truth is performed to make the ultrasound dataset beneficial. The software Pair (Shenzhen Duying Medical Technology Co., Ltd., Shenzen, China) is used to perform this step. According to suggestions of radiologists, the following points in the image annotation were abided: (1) In an ideal situation, the SP and fetal head are elliptical in the two-dimensional image; (2) In grayscale ultrasound images, the outer borders of the SP and fetal head mainly appear bright white. When there is no obvious white border, the boundary is determined according to the difference of the local gray value; and (3) The lower right corner of the pubic symphysis is adjacent to the bladder, and the boundary of the pubic symphysis should be determined with the bottom edge of the white area [1,[5], [6], [7], [8], [9], [10], [11], [12], [13]]. Following these points, all the segmentation labels were independently created with the software Pair by five students and visually reviewed independently by two radiologists to ensure accuracy. An example of mask images is shown in Fig. 3. Each was firstly loaded in the Pair software (Fig. 3A), images were then selected from videos in 10 frames each (Fig. 3B) and regions of the SP (Fig. 3C) and fetal head (Fig. 3D) were finally labeled.

Fig. 3

An example to help illustrate the label acquisition process using the Pair software. (A) Videos were loaded in the Pair software; (B)An image from a video was selected to label regions of the symphysis pubis and fetal head; (C) The symphysis pubis was labeled with red; (D) The fetal head was labeled with green.

Pair software directly generates ``*.tar'' files, and after decompression, ``*.json'' and ``*.nii'' files are generated. The ``*.json'' file records the coordinates of the key points on the contour, and the ``*.nii'' records the mask corresponding to the contour, whose value is the category id value set in the configuration file. Further, we set the SP pixel value in the mask to 7, and the pixel value of the Head to 8 to generate each label. Read the classification of different labels through ``FrameLabel'' in the ``*.json'' file (3: None, 4: OnlySP, 5: OnlyHead or 6: SPHead). Finally we convert the grayscale image to an RGB image (SP is red, Head is green). For the configuration parameters of the Pair software and the label reading code, we will provide the corresponding configuration files and codes in https://github.com/JNU-IFM/JNU-IFM-config-and-code.git.

Ethics Statements

This study was approved by the Medical Ethics Committee of NanFang Hospital of Southern Medical University (NFCE-2019–024).

CRediT authorship contribution statement

Yaosheng Lu: Conceptualization, Writing – original draft. Mengqiang Zhou: Investigation, Methodology, Visualization, Formal analysis, Writing – original draft. Dengjiang Zhi: Methodology, Visualization, Formal analysis, Writing – original draft. Minghong Zhou: Investigation, Methodology, Visualization, Formal analysis, Writing – original draft. Xiaosong Jiang: Methodology, Visualization, Formal analysis, Writing – original draft. Ruiyu Qiu: Investigation, Writing – original draft. Zhanhong Ou: Writing – original draft. Huijin Wang: Writing – original draft. Di Qiu: Data curation, Investigation, Writing – original draft. Mei Zhong: Investigation, Methodology, Visualization, Formal analysis, Writing – original draft. Xiaoxing Lu: Conceptualization, Writing – original draft. Gaowen Chen: Writing – original draft. Jieyun Bai: Conceptualization, Methodology, Visualization, Formal analysis, Writing – original draft.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Subject	Obstetrics, Midwifery and Women's Health
Specific subject area	The use of transperineal ultrasound techniques for the assessment of fetal head descent and progression is an adjunct to clinical examination.
Type of data	ImageTable
How the data were acquired	Data was obtained from transperineal ultrasound examinations that were performed in standard B-mode ultrasound using the Youkey D8 wireless 2D ultrasound probe with its corresponding supporting software (Wuhan Youkey Bio-Medical Electronics Co., Ltd., Wuhan, China). Data were labelled using the Pair software (Shenzhen Duying Medical Technology Co., Ltd., Shenzen, China) and then validated by two experienced radiologists.
Data format	RawAnalyzed
Parameters for data collection	The dataset includes intrapartum transperineal ultrasound (ITU) images and the corresponding segmentation labels of symphysis pubis (SP)-fetal head. In addition, four categories respectively corresponding to SP-fetal head images (SPHead) and other images (None: no SP and fetal head, OnlySP: no fetal head and OnlyHead: no SP) are included.
Description of data collection	The transducer was prepped by covering it with a surgical latex glove filled with coupling gel, then the prepped transducer, after applying gel, was placed between labia below the pubic symphysis to obtain a sagittal plane, small adjustments in the form of lateral movements of the probe were made until an image obtained showed clear maternal pelvic (pubic symphysis) and fetal (fetal skull) landmarks that did not show any shadows from the pubic rami. A total of 78 videos from 51 pregnant women were collected from NanFang Hospital of Southern Medical University between 2019 and early 2020. 6224 images are extracted from videos in 10 frames each. Images were labelled by using the software Pair.
Data source location	College of Information Science and Technology, Jinan University, Guangzhou, 510,632, China.
Data accessibility	Repository name: JNU-IFMData identification number: https://10.6084/m9.figshare.14371652
Related research article	Zhou M, Yuan C, Chen Z, et al. Automatic Angle of Progress Measurement of Intrapartum Transperineal Ultrasound Image with Deep Learning[C]//International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, Cham, 2020: 406–414.10.1007/978-3-030-59725-2_39