Datasets:
CoronaryDominance dataset
DATA DESCRIPTION The data we present here includes three parts - the main dataset, the real distribution dataset, and the domain shift dataset. P.S. IMPORTANT!!! The angiograms of the real distribution dataset are from the CoronarySYNTAX dataset.
The data we present here includes three parts - the main dataset, the real distribution dataset, and the domain shift dataset. The main dataset with 1,025 studies (319 of whom have left dominance, and 706 of whom have right dominance) from the Philips Allura Clarity cardiovascular imaging system is for model training and hyper-parameters fine-tuning. For the main dataset, each study falls into one of five categories: bad quality, artifact, small diameter, occlusion, or normal. ”Normal” is the largest category. We enhanced the main data set with additional left dominant cases and occlusions in order to address the problem of imbalance
The real distribution dataset consists of 400 randomly selected studies (54 left dominant, and 346 right dominant) for the Philips Allura Clarity cardiovascular imaging system. Due to this, unlike the main dataset, the real distribution dataset follows the same distribution as all the studies in the clinic. The real distribution dataset does not have group subdivision, so it includes studies with poor quality, artifacts, and high uncertainty. The domain shift dataset comprises 149 studies (52 left and 97 right dominant) from the Philips Azurion cardiovascular imaging system. This dataset also does not have group subdivision. However, we have enhanced it with an additional left dominant studies
The structure of the data is as follows: The root directory contains two subdirectories: ”Left Dominance” and ”Right Dominance”, which contain angiographic studies of left and right dominant patients, respectively. Each study consists of several angio-graphic views - gray-scale videos saved in Numpy’s compressed array format (.npz). The study directory also contains two subdirectories named ”RCA” (for the right coronary artery) and ”LCA” (for the left coronary artery), respectively.
The study folder follows the structure ”Study0xxxx study id”, where ”0xxxx” is the sequence number of the study and ”study id” is the study ID. The name of the .npz file containing the angiographic view video matches the series ID. The study ID and the series ID are unique within the dataset and follow the same format has the corresponding ID in DICOM (.dcm) files. Our data are fully anonymized; therefore the IDs in the dataset do not match those in the original .dcm source files. The sequence number is unique within a specific dataset. We have introduced this for the convenience of referencing, such as ”main dataset study 580” instead of ”1.1.11.11111.11.903846392438863822081872923423339322”. We extracted the primary and secondary angles from the fields (0018,1510) and (0018,1511) in the original .dcm file. You can consult the DICOM Standard Browser for more information on these angles. In most cases, the is collaterals tag is True when the is occlusion tag is also True.
‘‘Below is the data structure in .npz files.
{Name:”pixel_array”, Type:”Float array, size of frames×512×512”, Range: “0 - 255”, Description: “angiographic 3D view” }
{Name:”seriesid”, Type:”string”, Range: “consist of numbers and ‘.’ ”, Description: “unique ID of an angiographic view” }
{Name:”studyid”, Type:”string”, Range: “consist of numbers and ’.’ ”, Description: “unique ID of an angiographic study” }
{Name:”series_number”, Type:”integer”, Range: “0 - 1025”, Description: “ID of an angiographic view” }
{Name:”is_collaterals”, Type:”Boolean”, Range: “{True, False, None}”, Description: “Collaterals in LCA” }
{Name:”primary_positioner_angle”, Type:”Float ”, Range: “-90 - 90, None”, Description: “positioner angle, degree” }
{Name:”secondary_positioner_angle”, Type:”Float ”, Range: “-90 - 90, None”, Description: “positioner angle, degree”}
{Name:”is_occlusion”, Type:”Boolean”, Range: “ {True, False}, None”, Description: “Occlusion in RCA” }
{Name:”is_undefined”, Type:”Boolean”, Range: “ {True, False}, None”, Description: “Studies with high uncertainty” }
{Name:”is_artefact”, Type:”Boolean”, Range: “ {True, False}, None”, Description: “The presence of artifacts” }
{Name:”artery_type”, Type:”String”, Range: “{LCA, RCA}”, Description: “Coronary artery type” }
BACKGROUND INFORMATION AND TECHNICAL DETAILS ON THE DATASET
We release a new dataset containing invasive coronary angiograms for the coronary dominance classification task, an essential aspect in assessing the severity of coronary artery disease. The dataset holds 1,574 studies, including X-ray multi-view videos from two different interventional angiography systems, which allows one to test the effect of domain shift. Each study has the following tags: bad quality, artifact, high uncertainty, and occlusion. Those tags help to classify dominance classification more accurately and allow to utilize the dataset for uncertainty estimation and outlier detection.
Coronary dominance classification is an essential step in the SYNTAX (Synergy Between Percutaneous Coronary Intervention with Taxus and Cardiac Surgery) score estimation, which has become an crucial tool for assessing the severity of coronary artery disease. This classification is also necessary when determining the severity of coronary lesions using the Gensini score. Coronary dominance is determined by the coronary artery branch that supplies the posterior descending artery (PDA). There are three main categories: left dominance, right dominance, and co-dominance [3]. Approximately 70-80% of the population has right coronary dominance, while approximately 5-10% have left coronary dominance. In some regions, the proportion of individuals with left coronary dominance may be as high as 20%. Co-dominance occurs in approximately 10-20% ofcases, where the PDA receives blood from the left and right coronary arteries. It is important to note that co-dominance is not in the SYNTAX scoring system. Therefore, when dealing with complex cases involving arteries supplying the PDA almost equally, clinicians have to choose ’right dominance’ when calculating SYNTAX scores.
Interventional cardiologists use X-ray video, known as angiographic view, to estimate the SYNTAX score and classify coronary dominance. To get these videos,interventional cardiologists inject a contrast agent into the left (LCA) and right (RCA) coronary arteries. By capturing heart pulses from different angles, an angiography study provides valuable information about the cardiovascular system. The following factors may complicate the classification of coronary dominance: a total occlusion, a small RCA diameter, a poor-quality angiogram, and artifacts. A RCA with a small diameter means there is significant disagreement among experts about how to classify a patient. Cases with a poor-quality angiogram, which can lead to incorrect diagnoses, typically have low-contrast medium filling. Studies that involve artifacts include those with pacing electrodes or sternal wires. Our dataset includes all these categories and allows us to measure the impact of each factor on classification metrics.
The domain shift is an important problem when working with medical data. A change in a medical device's parameters or settings could significantly impact the neural network's prediction accuracy. To allow ML researchers to experience this phenomenon, we present angiograms from two different cardiovascular imaging systems in our dataset - Phillips Allura Clarity and Philips Azurion.
Coronary angiography is an essential tool for medical imaging that allows for real-time observation of the heart and blood vessels in a 3D space. Using X-ray technology, this technique captures images on 2D planes and creates video images of the internal organs and blood vessels. These views are not synchronized in time. The study consists of the LCA and RCA views, which typically contain between 20 and 70 frames, 512×512 each. The pixels values are in a range from 0 to 255. The angiograms from Phillips Azurion have a frame around the image which we deleted the frame and resized the images to 512×512.
Six interventional cardiologists with experience ranging from 1 to 25 years provided ground truth information on coronary dominance. A moderator PhD-level cardiologist with more than 15 years of experience verified their labels. We considered his labels to be a gold standard. He also labeled studies with additional tags, such as occlusions and artifacts. The disagreement rate between the experts was 2.8%. After excluding complex categories such as poor quality, artefacts, and small RCA diameters, the disagreement rate on the main data set dropped to 1.5% Our dataset of angiographic studies for coronary dominance classification could be a valuable benchmark for 3D multi-view methods. Since the data come from different cardiovascular imaging systems, one can test the effect of domain shift. The dataset is useful for estimating uncertainty and detecting outliers. If one trains a neural network using data without artifacts and then tests the network on a subset with artifacts, one can estimate the algorithm’s stability concerning outliers. Studies involving RCA with a small diameter are particularly interesting for uncertainty estimation since there is significant disagreement among experts regarding dominance classification in these studies.
We recommend using normal and occlusion studies from the main dataset for model training and the other studies for testing only. For normal studies, one could use both the RCA and LCA views for coronary dominance classification. However, in our opinion, the RCA views are more suitable for this purpose. For studies with occlusion, only LCA views provide essential information for the dominance classification.