ViewPosition information
Hi, I was looking through the metadata json files and found that out that 87% of the images have an empty string for ImageViewPosition. Is there information on what view position these images actually are? I saw in the paper that
For radiographic view types, we observed similar distributions across splits, with lateral (LAT) views being
most common (approximately 35%), followed by posteroanterior (PA) views (approximately 31%)
and anteroposterior (AP) views (approximately 29%)
Were these statistics only for the images with valid dicom metadata or determined with some other method (on the whole dataset)?
Hi, thank you for bringing this up.
We're currently checking the view position metadata for those images with empty strings in the ImageViewPosition field, and will update the repo once we finish.
Any chance that we can get the study time as well? Its hard to determine the order of a patient's studies with just the study date, e.g.:
df[df['subject_id'] == 'pGRDNZHPFEECMZ3IE'][['subject_id', 'StudyInstanceUid', 'StudyDate']]
subject_id StudyInstanceUid StudyDate
1556 pGRDNZHPFEECMZ3IE 1.2.826.0.1.3680043.8.498.81416886309946122883... 2013-02-16
4163 pGRDNZHPFEECMZ3IE 1.2.826.0.1.3680043.8.498.97027067852823946045... 2013-02-16
8052 pGRDNZHPFEECMZ3IE 1.2.826.0.1.3680043.8.498.45953479062477953058... 2013-02-22
Hi, I was trying to map the images to their prior image in cases where "Comparison" is given but its getting confusing to locate the prior image form just the string like:
"Comparison with prior exam from 04/10/2015." as there are multiple records with same date. Is there a way we can map these images to their priors
@XiaomanZhang
?
Appreciate the work you and your collaborators have done @XiaomanZhang . Is there any update on this (at least for the ViewPosition)?
Thanks for following up! I've updated the data with view position information extracted from the metadata study description fields in the DICOM files. This helped fill some of the empty ImageViewPosition. However, there are still some images missing this information after the metadata extraction. For these remaining cases, I used GPT-4o to classify them as AP/PA/LATERAL/UNKNOWN based on image content. After these, approximately 34k images still remain with "UNKNOWN" view positions. If you have any suggestions or better methods to determine view positions for these remaining images, I'd welcome the input!
You could evaluate the accuracy of the view predictions on a test set of images with known view positions.
Hi @XiaomanZhang , you can use the HealthGPT model (https://github.com/DCDmllm/HealthGPT) to assign ViewPosition labels for the remaining images. I have used it before and it is pretty straightforward