Known limitations

This page lists what the app does not do. Where a limitation is a decision, the reason is given.

Burnt-in text detection is partial

The check before anonymising marks areas that look like burnt-in text: very bright or very dark marks with sharp edges that stay in the same place while the anatomy changes. This finds patient banners and hospital names on a dark background.

It does not find small print, text over anatomy, low-contrast text, or text on images it did not examine. It examines two images per series. It does not read the text it finds.

The check can only add warnings. If it reports nothing for a series, nothing was noticed in the two images it examined; the series can still contain text. The app never reports a series as free of burnt-in text. You need to check every image yourself.

Text recognition (OCR) is not planned. It would make the check take seconds to tens of seconds per study, and it is least reliable on the text this check misses: small text, low-contrast text and text over anatomy.

Erasing or cropping a compressed image makes the file larger

Pixels cannot be changed inside compressed image data. When a compressed image has to change, because of an erased area, a crop, or a multiframe run being split into single images, the app decodes it and writes it uncompressed (explicit VR little endian). The tags that describe the pixels are updated to match the decoded data.

The uncompressed file is larger: the JPEG test image in the repository grows from 49 kB to 768 kB. A compressed image that does not need to change is uploaded unchanged. Setting the window does not change the pixels, so it does not count as a change.

DICOM video is an exception: its frames are uploaded as JPEG (see below).

Some images cannot be erased, cropped or split

Erasing, cropping and splitting multiframe runs work on images with 8-bit or 16-bit samples and with the colour values of each pixel stored together. The app refuses to erase, crop or split images stored in other ways, such as 32-bit samples, 1-bit segmentations or subsampled colour (YBR_FULL_422). Such files are counted among the files that could not be anonymised and are not uploaded. If they need none of these changes, they are uploaded as they are. These formats are rare in clinical studies.

DICOM video is uploaded as frames

A run stored as video (MPEG-2, H.264 or HEVC), as ultrasound and angiography systems often store them, is decoded into frames and uploaded as a series of single images. This is needed because Radiopaedia does not display video in DICOM files, and because erasing burnt-in text requires individual frames.

• Size. Each frame is uploaded as a JPEG at quality 95, after erasing and cropping. A 300-frame clip at 1024×768 is about 28 MB, compared with about 700 MB uncompressed. This is a second lossy compression of frames that were already lossy; the difference is not visible. Images that were lossless are never compressed this way. While the session is open, the decoded clip is stored uncompressed on disk: about 700 MB for the same clip. If decoding would leave less than about 500 MB of free disk space, the app refuses before it starts and shows how much space it needs.
• Time. Compressing the frames takes most of the anonymisation time: about 18 seconds for 300 frames. The first time you open a video, the whole video is decoded, which takes a few seconds for a long clip; after that, frames appear immediately.
• Only the images. Audio, the second view of a stereo clip and the frame rate are not uploaded.
• Frame count. If the video contains fewer frames than its header states, the frames it contains are uploaded and the missing ones are reported.

Video is decoded with ffmpeg, which is included with the app. The Info panel shows the ffmpeg version, or "no video decoder" if the build has none that works.

Formats the app cannot read

The app cannot decode a few rare transfer syntaxes, such as JPIP (where the pixel data is on a server) and deflated DICOM. A multiframe run in such a format cannot be selected: its card names the format, and the number of such stacks is shown next to the number of selected stacks. A single image in such a format is uploaded unchanged, but it cannot be erased or cropped.

No curved reformats

Reformat creates images in the patient's axial, coronal and sagittal planes, and in oblique planes by rotating the crosshair. It cannot create a curved reformat along a vessel or a nerve. Curved reformats are not planned, because a case usually needs only a standard reformat to show a finding in another plane.

If the files do not contain ImageOrientationPatient, the app uses the planes of the acquisition, which are correct only for an axial acquisition. The dialog shows a note when this happens.

Enhanced MR and CT cannot be reformatted

A dynamic enhanced MR or CT file is split into its phases, and each phase can be uploaded. Reformat cannot build a volume from it, because an enhanced file stores the pixel size per frame, which the app does not read for reformatting. The dialog shows "These images have no pixel spacing (PixelSpacing), so a reformat would have no scale".