Data Information and Details
Two databases will be used in the challenge: Abdominal CT and MRI. Each data set in these two databases corresponds to a series of DICOM images belonging to a single patient. The data sets are collected retrospectively and randomly from the PACS of DEU Hospital. There is no connection between the data sets obtained from CT and MR databases (i.e. they are acquired from different patients and not registered).
Data Quality and Specifications:
The first database contains CT images of
30 40 different patients. These patients are potential liver donors, who have healthy (no tumors, lesions or any other diseases) liver. The CT images were acquired from upper abdomen area of the patients at portal venous phase after contrast agent injection. Portal venous phase is the phase obtained 70-80 sec after contrast agent injection or 50-60 sec after bolustracking. In this phase the liver parenchyma enhances maximally through bloodsupply by the portal vein. Portal veins are well enhanced and some enhancement of the hepatic veins is also seen on the portal venous phase. Therefore, this phase is widely used for the liver and vessel segmentation prior to surgery.
Three different modalities, Philips SecuraCT with 16 detectors and a Philips Mx8000 CT with 64 detectors and Toshiba AquilionOne with 320 detectors (all equipped with the spiral CT option), are used. The patient orientation and alignment is the same for all data sets. Each data set consists of 16 bit DICOM images with a resolution of 512x512, x-y spacing between 0.7-0.8 mm and having 3 to 3.2 mm inter-slice distance (ISD) (i.e. smaller ISD is not being used for these acquisitions due to the routine clinical procedure). This corresponds to an average of 90 slices per data set (i.e. minimum 77, maximum 105 slices). In total, 1367 slices will be provided for training and 1408 slices will be used for tests. The challenges of the data set can be summarized as:
2. Varying Hounsfield ranges for the same tissue across data sets due to the contrast media
3. Significant shape differences of anatomical structures across patients
4.15% of the database contains atypical liver shapes (i.e. unusual size or orientation of the liver).
Figure 1. Example images from database 1 (Abdominal CT). Slices show (a) very low contrast difference and unclear boundary between the heart and the liver; (b) unclear boundary due to partial volume effects between the right kidney and the liver; (c) contrast enhanced vascular tissues inside the liver parenchyma; (e) relatively less enhanced vessels compared to (c).
While developing and training the algorithms, it is possible to include other liver datasets such as SLIVER or personal institution data along with the datasets obtained from this challenge. We ask the participants to note this information in their entries. Besides using other data sets, utilization of new approaches such as transfer learning to fine tune a trained model to abdominal organ segmentation or alternative strategies such as data augmentation are encouraged to reflect the effects of the latest developments on the field.
The second database includes
90 120 DICOM data sets from two different MRI sequences (T1-DUAL in phase ( 30 40 data sets), out phase ( 30 40 data sets) and T2-SPIR ( 30 40 data sets)), each of which is being routinely performed to scan abdomen using different radiofrequency pulse and gradient combinations. This database also does not include any tumors or lesions at the borders of the annotated organs of interest (i.e. liver, kidneys, spleen). The data sets are acquired by a 1.5T Philips MRI, which produces 12 bit DICOM images having a resolution of 256 x 256. The ISDs vary between 5.5-9 mm (average 7.84 mm), x-y spacing is between 1.36 - 1.89 mm (average 1.61 mm) and the number of slices is between 26 and 50 (average 36). In total, 1594 slices (532 slice per sequence) will be provided for training and 1537 slices will be used for the tests.
SPIR (Spectral Pre-Saturation Inversion Recovery) stands for a hybrid imaging sequence and uses T2-weighted contrast mechanism. For selective suppression of fat protons, the pre-saturation pulse is applied separately to each slice selection gradient. This sequence requires sensitive adjustment of calibration and a very homogenous magnetic field. The above mentioned features of SPIR makes it a preferred sequence to study liver, because the liver parenchyma can be analyzed very well with suppression of the fat content inside the parenchyma. Moreover, the abdominal organs’ border appearances get visually clearer, because of the suppression of the fat tissue around them. Being T2-weighted, it is also possible to navigate the vessels within liver as they appear hyper-intense. The adjacent abdominal organs and tissues become more separable from each other with their high signal values. One more important contribution of the SPIR sequence is its low sensitivity to motion. This feature provides minimization of the artifacts that adversely affects image quality in abdominal studies.
Figure 2. Samples of abdominal MRI images from T2-SPIR sequence
T1-DUAL (in-phase and out-phase) is a fat suppression sequence, which uses the difference in T1 times of fat and water protons. The signal is acquired twice: first when water and fat protons are in phase and second, when they are out of phase (while exciting protons are returning to their first position). For 1.5 Tesla devices the in-phase time, which water and fat protons are in same direction is 4.6 milliseconds and the out-phase time, which fat and water protons are opposite directions, is 2.3 milliseconds. By determining TE (Time of Echo) value with this information, fat suppression is accomplished by subtracting corresponding frequencies of fat and water signals. This sequence is very useful to understand the fat content in lesions. Since T1-DUAL is a T1-weighted sequence, it is very effective to identify blood and tissues that are rich in protein. This sequence also helps determining the level of liver lubrication. In out-phase images, the border of the organs appears to be black, due to the sudden change in the amount of fat and water at the organ boundaries that cancels the acquired signal. This property of T1-DUAL is sometimes used for border delineation algorithms.
Figure 3. Samples of abdominal MRI images from T1-DUAL (in-phase) sequence
Training and Testing Data
The training data of both databases will be distributed before the challenge. The data will be available after online registration of participants and signing a letter of intent. The training data contains complete series of images and their ground truths for the selected data sets (i.e. patients). In order to provide sufficient data that contains enough variability to be representative of the problem, the data sets in the training data are selected to represent both the difficulties that are observed on the whole database (e.g. partial volume effects for CT or bias fields for MRI) and examples of the rare but important challenges such as atypical liver shapes (Figures 4 and 5). It is planned to share 50% of whole data sets (i.e. 15 data sets per database). It will be allowed to use additional data sets (for instance the data provided at SLIVER07 challenge) for training, if the teams needed more data to train and/or prepare their system (Please note that the livers in SLIVER data sets contains several pathologies and various patient orientations).
Figure 4. Examples of challenges from the training data of the first database (abdominal CT images of liver transplantation donor candidates) (a) Unclear boundary between the liver and the heart. (b) Liver has three dis-connected components on a single slice (c) Atypical liver shape, which causes unclear boundary with spleen (d) Varying Hounsfield range and non-homogeneous parenchyma texture of liver due to the injection of contrast media.
Figure 5. Examples of challenges from the training data of the second database (abdominal MRI) (a) sudden changes in planar view and unclear boundary (spleen-left kidney). Effect of bias field in (b) T1-DUAL, and (c) T2-SPIR.
Remaining 50% of the whole data sets will be used as test data. The test data will be distributed at the day of the challenge. The ground truths of the test part will never be given to the participants. The participants will submit their results as binary or labeled data. The evaluation will be performed on-site and immediately after the submission of a result.