Label Studio Adds DICOM Interface That Skips PNG Conversion

Why HumanSignal wants to avoid PNG exports

HumanSignal argues that most medical imaging teams discard the file format before labeling begins. A study is run through a conversion script and ends up as a folder of PNGs. According to the company, what gets lost includes the windowing ranges set by the technician, the structures the file already encodes, the geometry linking one slice to the next, and the acquisition metadata.

Annotators then work on flat images in a general-purpose image tool and have to rebuild the volumetric context mentally. HumanSignal says its new interface avoids this by reading the DICOM file itself.

Two views, windowing and pre-existing labels

The interface reconstructs the study from the DICOM and shows two views at once: a frontal reconstruction and a side reconstruction. A slider moves through the volume, and both views track the same position. Zoom, pan and reset controls are included. In an exported image stack, HumanSignal notes, an annotator cannot tell where a slice sits in the volume or what lies above and below it.

Windowing is handled as a control rather than being fixed in an export. Users can pick from several options:

• the header defaults, meaning the ranges set by whoever authored the DICOM set
• the full unfiltered range as stored in the file
• standard presets for bone and soft tissue
• a dental view that isolates the teeth

Switching takes one click and changes nothing on disk, because it is only a rendering choice. HumanSignal points out that a window tuned for bone renders soft tissue as flat grey, while a soft-tissue window loses skeletal detail.

Many DICOM files already carry labeled structures, which the company says most pipelines discard during conversion. The new interface reads those labels from the file, so each encoded structure arrives segmented across every slice. Annotators can then correct and refine masks with a brush and an eraser instead of starting from a blank image. An overlay-strength control sets the opacity of the label colors so a boundary can be checked against raw pixels. HumanSignal says this shifts the work from production to review.

3D rendering and header metadata

The interface renders the full volume in 3D with WebGL. The 3D and slice views share state: selecting a structure in the 3D render and returning to the slice view keeps that structure selected at its position in the volume. According to HumanSignal, this removes the slowest part of volumetric annotation, which is finding the slice where the target lives.

A sidebar beside the scan lists the DICOM header contents, including acquisition parameters, spacing, orientation, equipment and the ranges set by the original author. The company expects annotators will not read it often, but says it matters when a scan looks off-scale, mis-oriented or out of step with the rest of a batch. A glance can settle the question, rather than an annotator flagging the task and waiting two days.

What it means for annotation teams

HumanSignal says that reading DICOM instead of converting it lets annotators work in the views their domain uses, start from structures the file already encodes, and check questions against the file's own metadata. In the company's words, the team spends its time on judgment calls rather than rebuilding context the format carried all along.

The interface is configurable. According to HumanSignal, the windowing presets, the tool set, the layout and which structures stay editable are all defined in code. It runs on Label Studio Enterprise's programmable interfaces, so teams can adapt it to their own modality and workflow. The source does not give pricing or availability details.