Redactwizard DICOM Research – User Guide

Version 1.1

Redactwizard DICOM Research is a portable desktop workstation for the scientific viewing and processing of medical imaging data. The software supports DICOM de-identification, pixel-level redaction of personally identifiable information, measurements, annotations, and experimental multimodal AI image analysis. In addition to DICOM datasets, standard image formats can also be processed.

The application supports accessibility features such as screen-reader-readable interface labels and tooltips. Accessibility is continuously tested and improved, including keyboard navigation and support for assistive technologies.

Local data processing

Redactwizard processes image data and runs its integrated AI analyses locally on your computer. The integrated AI functionality does not require cloud-based inference. The application is designed to process data locally without transferring image data to external AI or cloud services.

Notice

Redactwizard DICOM Research is a research and assistance tool, not a certified medical device. Neither the software nor its integrated AI models provides guaranteed diagnoses, clinical findings, or treatment recommendations. Automated recognition functions may overlook, misinterpret, or incompletely process relevant or sensitive information. All clinical assessments and decisions must be made by qualified medical professionals.

1. Starting the portable version

Redactwizard is distributed as a portable application. A conventional Windows installation is not required.

Extract the complete program folder to a location of your choice, such as a fast internal NVMe SSD. Start the application using the executable file Redactwizard.exe.

The executable must remain in its designated directory together with the associated program files and the config.json configuration file. Do not move or delete individual components of the program folder independently.

First-launch loading time

Initial initialization

During the first program launch and the first loading of the Vision-Language Model, initialization may take up to one minute, depending on processor performance, disk speed, and GPU hardware. During this process, the model weights and vision projector are verified in the model directory and loaded into system memory (RAM) and/or graphics memory (VRAM). The required tensors are then initialized.

Please do not close Redactwizard while this process is running.

Windows security notice: Microsoft Defender SmartScreen

When the application is launched for the first time, Windows may display the Microsoft Defender SmartScreen security screen. This can occur with specialized applications that are not distributed through the Microsoft Store or for which Windows has not established sufficient reputation information.

Verify the application before running it

A SmartScreen warning does not prove that a file is malicious, nor does the absence of a warning guarantee that it is safe. Before proceeding, check the application's origin and make sure you obtained the file from a trusted source. Do not bypass the security warning if you cannot establish the file's origin or integrity.

Step 1: If you obtained the application from a trusted source and have verified the file, click More info if necessary.

Step 2: Review the displayed application details. If you have verified the file's origin and explicitly wish to allow it to run, you can proceed by clicking Run anyway. The exact wording may vary depending on your Windows version.

2. Windows reports a missing Visual C++ component

Redactwizard requires the appropriate C++ runtime components for certain native libraries. If the application fails to start because MSVCP140.dll, VCRUNTIME140.dll, or a comparable runtime file is missing, installation of the official Microsoft package may be required.

Use the official Microsoft package

Download the Visual C++ Redistributable for Windows x64 exclusively through the official Microsoft link:

Download Microsoft Visual C++ Redistributable x64

Run the downloaded installer and follow the instructions. Then restart Redactwizard. If Windows requests a system restart, restart your computer as well.

Do not download individual DLL files from unofficial download websites. Such files may be outdated, incompatible, or tampered with.

3. Licensing, language, and workspaces

License activation

Redactwizard requires a valid license key to unlock its protected file-handling and analysis functions.

• Enter the license key: Click the key icon in the upper-right corner of the menu bar. Paste your license key into the input field using CTRL + V, then confirm by clicking Activate.
• Clipboard: Use CTRL + C to copy a previously selected license key.
• Storage: The plaintext license key is not stored on disk. Instead, a cryptographic hash is stored in the config.json file.
• Restricted operation without a license: Without an activated license, the protected functions for opening DICOM files, loading series, and running AI analyses cannot be used.

Important licensing information

Keep your license key secure. Do not share it publicly or enter it anywhere other than the designated Redactwizard license activation interface.

Language, theme, and tab system

• Language selection: Use the globe/translation icon in the upper-right corner to switch between German and English.
• Color mode: Use the sun/moon icon to switch between the dark appearance (Dark Mode) and the light appearance (Light Mode).
• Multiple workspaces: Click the plus icon (+) in the upper tab bar to open another workspace. Each tab manages its own image series and associated viewing and working data, such as zoom level, measurements, and cache data.

4. AI model settings

Click the gear icon in the menu bar to access the performance settings of the locally executed Vision-Language Model.

These settings affect how the model uses memory, computing resources, and output length.

• Context Size: Defines how many tokens the model can consider within its context. The default value is 8192 tokens. Higher values allow longer contexts but may require substantially more RAM or VRAM.
• Batch Size: Determines how many units are processed together during certain processing operations. The default value is 512.
• Max Output Tokens: Limits the length of the generated response. The default value is 512 tokens.
• GPU Layers (n_gpu_layers): Specifies how many model layers are offloaded to the graphics card. A value of 0 means CPU operation without offloading model layers to the GPU. The appropriate value depends on the model, available memory, and the backend in use.
• Vision Projector on GPU (mmproj GPU): Enables GPU acceleration for visual encoding, if supported.

Tip: Optimize GPU settings gradually

If you have a dedicated GPU, you can increase the GPU layer count gradually while monitoring utilization and available VRAM. On a GPU with 4–8 GB of VRAM, the achievable value may vary depending on the model and other settings. A higher value is not automatically the best setting for every system.

Restart required

Changes to GPU and model parameters may only take effect after restarting Redactwizard.

OCR Text Recognition

Max. Image Edge Length (Resolution)

Controls the maximum length in pixels of an image's longest edge prior to OCR processing. Larger pages are proportionally downscaled.

• Lower values (800–1,100 px): Result in faster processing and lower memory usage.
• Higher values (1,500–2,000 px): Preserve finer lines and details. Recommended for documents with small fonts, signatures, or handwritten notes to significantly increase recognition accuracy.

OCR CPU Threads

Specifies how many CPU cores are used in parallel for image analysis and text recognition. A higher value speeds up the processing of multi-page documents, while a lower value preserves system resources for other tasks.

5. DICOM header de-identification

DICOM files can contain numerous personal, medical, and institutional details in their headers. Redactwizard provides de-identification functions designed with relevant DICOM requirements in mind.

Click the broom icon (BROOM) in the toolbar to open the de-identification dialog. There you can select which data categories should be processed.

1. Patient identifiers: Patient names are replaced with ANONYMIZED. Other details, such as patient ID, accession number, date of birth, address, and telephone numbers, are hashed or removed according to the implemented de-identification function.
2. Treating physicians and institutions: Institution names, addresses, and details about referring and performing physicians, such as Referring Physician, Performing Physician, and Operators Name, can be removed.
3. Free-text comments: Patient, study, and image comments may contain personally identifiable information. Redactwizard processes, among others, tags 0010,4000, 0028,4000, and 0032,4000.
4. UIDs and dates: Study, series, and SOP Instance UIDs, as well as date values, can be pseudonymized. When pseudonymization is performed consistently, associations between related series are intended to be preserved within the processed dataset.
5. Private tags: Manufacturer-specific and proprietary metadata can be removed if the relevant option is selected or the data is covered by the implemented function.

Confirm your selection using Save anonymized copy.... For a single file, choose a destination filename. For complete series, choose a destination folder in which the processed files will be saved.

Important: De-identification does not guarantee complete anonymization

Individual identifiers, proprietary metadata, free-text fields, burned-in image text, or other identifying features may remain after processing. Pseudonymization is also not automatically equivalent to complete anonymization.

Independently review the processed DICOM files and their associated image content before sharing, publishing, or otherwise disclosing them. In particular, verify that any remaining information is appropriate for the intended purpose.

6. Image redaction and AI-assisted text recognition (OCR)

De-identifying the DICOM header does not automatically remove personal information burned directly into the image pixels. Such information may appear, for example, in ultrasound images, scanned documents, or older X-ray images.

Warning for burned-in image information

When opening a DICOM file, Redactwizard checks the BurnedInAnnotation (0028,0301) tag. If its value is YES, the application displays a warning bar at the top of the image.

Review burned-in information

The value BurnedInAnnotation (0028,0301) can indicate the presence of burned-in information. A different value or a missing entry does not guarantee that an image is free of personally identifiable information. Therefore, inspect the image pixels themselves as well.

Manual redaction with the pencil tool

1. Click the pencil icon (PENCIL_SIMPLE) in the toolbar. The active tool is highlighted in blue, and a crosshair appears over the image.
2. Hold down the left mouse button and drag a rectangle over the text or image area to be redacted.
3. After releasing the mouse button, the selected area appears as a red-outlined box. You can add as many additional areas as needed.

Automatic OCR text recognition

1. Click the bounding-box icon (BOUNDING_BOX).
2. A loading indicator appears while Redactwizard processes the image. The integrated text recognition engine analyzes the image for potential text structures.
3. Recognized text regions are displayed as orange frames over the image.
4. A plausibility filter is intended to prevent anatomical structures from being incorrectly classified as text.

Always review OCR results

Automatic text recognition may miss text, mark incomplete regions, or misclassify image structures. Manually inspect all suggested redaction areas and add any missing regions as necessary.

Applying or correcting redactions

• Delete an individual box: Right-click a red or orange box to remove it.
• Discard suggestions: Use the red trash icon (TRASH) to discard boxes that have not yet been permanently applied to the image pixels.
• Apply to an entire series: For multi-image series, activate the Apply to entire series checkbox to transfer the masking regions to additional slices.
• Permanently apply redaction: Click the green checkmark icon (CHECK) to permanently set the selected pixels to [0, 0, 0]. The redaction is embedded in the image pixel buffer and is carried over to supported export formats.

Caution: Permanent pixel redaction

Permanently overwriting image pixels can irreversibly alter the original image information. Carefully review all marked regions before applying the redaction, and keep an unchanged copy of the original file in a secure location when necessary. Also inspect the exported file before sharing or publishing it.

7. Toolbar, image controls, and annotations

The top toolbar combines file management, image navigation, transformations, measurements, annotations, and export functions.

Interactive buttons provide tooltips and visually indicate the active state of the selected tools.

File management

• Open file (FILE): Opens individual DICOM files (.dcm), PNG and JPEG images, and supported .zip archives.
• Open ZIP archives: When you select a ZIP archive, for example a PACS web download, Redactwizard extracts its contents in the background into a temporary cache and loads recognized slices as a related series.
• Open DICOM series (FOLDER_OPEN): Opens a folder selection dialog for directories containing uncompressed DICOM files.

Check ZIP archives before processing

Only open archives from trusted sources. Make sure sufficient disk space is available and that the contained files are suitable for processing.

Image navigation and transformation

• Pan (HAND): Move the image within the viewer by dragging it with the mouse button held down.
• Windowing / Contrast (CIRCLE_HALF): Adjust Window and Level by dragging with the left mouse button. Horizontal movement changes the window width and contrast, while vertical movement changes the window center and brightness.
• Fit to window (ARROWS_OUT): Scales the image to fit within the available viewer area.
• Original size (CORNERS_IN): Switches to a 1:1 pixel view.
• Zoom: Enlarge or reduce the display using the mouse wheel.
• Rotate (ARROW_CLOCKWISE): Rotates the displayed image clockwise by 90 degrees.
• Flip (FLIP_HORIZONTAL): Flips the view horizontally.
• Reset view (ARROW_COUNTER_CLOCKWISE): Restores the default viewing geometry.
• Previous image (SKIP_BACK) and Next image (SKIP_FORWARD): Navigate through the slices in the loaded series. You can also use the image slider. The Image X/Y display indicates the current position within the series.

Measurement and annotation tools

Active tools are indicated by a blue highlight.

• Ruler / Distance measurement (RULER): Draw a measurement line across a structure. Redactwizard calculates the distance in pixels or, when suitable PixelSpacing values are available, in millimeters.
• Arrow pointer (ARROW_UP_RIGHT): Places a directional arrow to highlight a specific image region.
• ROI rectangle (RECTANGLE): Draws a rectangular region of interest. The application displays its width and height, for example 120 × 80 px.
• Finding note (TEXT_T): Places a note marker on the image. In the German interface, notes are numbered Befund 1, Befund 2, and so on. In the English interface, they appear as Finding 1, Finding 2, and so on.
• Delete all (TRASH): Removes the annotations, rulers, and markers managed by the corresponding tool area.
• Remove an individual annotation: Right-click an existing marker to delete it.

Interpreting measurement values

The accuracy of a distance measurement depends, among other things, on the available image metadata and image geometry. If suitable PixelSpacing values are missing or unreliable, measurements displayed in millimeters may not be reliable. Verify the relevant conditions before documenting measurements for research purposes.

Export functions

PNG snapshot (CAMERA): Opens a file dialog and saves the current image as a PNG. Depending on the active functions, displayed annotations, measurement values, arrows, finding markers, and redactions are included in the exported image.

Inspect the exported file to ensure that all intended markers and redactions are present and that no unwanted personally identifiable information remains visible.

8. Multimodal AI image analysis and DICOM metadata

Redactwizard displays the image and analysis information in separate interface elements. This keeps the image area as large as possible while analysis results and metadata are displayed in floating windows.

Floating AI analysis window (SPARKLE)

Click the sparkle icon (SPARKLE) in the toolbar, to the left of the broom icon, to display the AI analysis window.

• Position and size: You can move the window within the workspace, resize it, or collapse it.
• Preset prompts: Choose from the available analysis instructions:
  • Describe image
  • Describe visible findings
  • Systematic image description
  • Identify unusual features
  • Objective image assessment
• Custom prompt: Enter an individual instruction in the multiline text field.

Write custom prompts in English

The Vision-Language Model was primarily trained on English-language medical image-text pairs. Therefore, formulate custom analysis instructions in English to match the model's intended input format.

Example: Examine the apical lung zones for pneumothorax or consolidation

• Start analysis: Click Analyze current image (SPARKLE). Generation progress is displayed through a token counter.
• Cancel analysis: While an analysis is running, a cancellation button is available. Clicking Cancel analysis stops the ongoing token generation, provided the model engine supports cancellation at that point.
• Copy results: You can copy the generated text for further documentation.
• Export PDF report: The results can be converted into a formatted, multi-page A4 report through the export function. The report may include the image, generated text, a disclaimer, and a cryptographic license hash.

AI outputs are not verified medical findings

Model-generated text may sound plausible while still containing errors. Review the output against the underlying image data and do not use it as an independent medical assessment or as the basis for clinical decisions.

Floating DICOM metadata window (IDENTIFICATION_BADGE)

Click the identification badge icon (IDENTIFICATION_BADGE) to display the transparent metadata window in the upper-right corner.

Depending on the available information, it may display the following details:

• Study: Modality, such as CT, MR, DX, CR, or US; body part; series description; and acquisition date.
• Patient: Patient name or anonymized name, patient age, and sex.
• Technical details and geometry: Slice thickness, pixel spacing, tube voltage (kVp) for CT or magnetic field strength for MRI, scanner manufacturer, and Window/Level values.

The displayed information is derived from the available DICOM metadata. Missing, incomplete, or incorrect metadata may result in incomplete or misleading displays.

9. Targeted analysis and annotations

Targeted prompting with annotations

You can use annotations to direct the AI analysis toward a specific image region.

1. Mark the desired area using the arrow (ARROW_UP_RIGHT) or add a finding note, for example Finding 1.
2. In the AI window, formulate an English instruction referring to the selected annotation.

3. Example prompt:

   Focus on the region indicated by the arrow / Finding 1. What image features are visible?

The annotation serves as a reference for the task. Whether and how the model actually takes the marker into account depends on the implementation and input data. A targeted marker therefore does not guarantee a correct interpretation of the selected region.

Tip: Ask specific questions

Formulate prompts as clearly as possible and focus on the question you want to investigate. Always review the response against the image data and remember that the model may overlook or misdescribe features.

10. Probabilistic nature and limitations of the AI model

Redactwizard DICOM Research uses a locally executed Vision-Language Model that generates text based on the supplied image data and the corresponding input. Its outputs are based on statistical model computations and must not be equated with a verified medical assessment.

Among other things, the model may:

• Misinterpret visible image features.
• Describe features that are not clearly present in the image data.
• Overlook relevant abnormalities.
• Generate incomplete or misleading responses.
• Produce different results depending on image quality, imaging modality, and prompt formulation.

Even linguistically convincing results may be factually incorrect. Correct, complete, or reproducible outputs are not guaranteed.

An abnormality not mentioned by the model is not ruled out

If an AI-generated response does not mention an abnormality, this does not mean that a disease or injury can be ruled out. Likewise, the model's mention of a feature is not reliable proof that a particular disease is present.

The results are intended exclusively for experimental, AI-assisted exploration of image data. They do not replace a qualified medical examination or assessment.

11. Intended purpose, usage restrictions, and liability notice

Redactwizard DICOM Research is intended for the technical viewing and processing of medical imaging data, as well as for research, teaching, training, and software evaluation. Its AI functions are intended for experimental, AI-assisted exploration of image data.

The software is not intended for clinical diagnosis, the generation of clinically usable reports, or supporting decisions concerning the prevention, diagnosis, treatment, or monitoring of patients.

Redactwizard DICOM Research and its AI-generated outputs are not substitutes for a qualified medical examination or assessment. AI outputs must not be used as medical diagnoses, radiological findings, or treatment recommendations. They must not be used as the basis for clinical decisions or patient care measures.

Limitations of de-identification and redaction

The integrated functions for detecting, de-identifying, and redacting personally identifiable information may be incomplete. In particular, identifying information in DICOM metadata, burned-in image text, or other image components may remain undetected.

Complete anonymization or the removal of all personally identifiable information is not guaranteed. Before sharing, publishing, or otherwise disclosing processed files, independently review their metadata and image content.

Data protection and lawful use

Users are responsible for determining whether their processing of personal data and any subsequent disclosure or publication are lawful. Personal health data may only be processed where an appropriate legal basis exists and the applicable data protection and security requirements are met.

Responsibility when processing medical data

Before processing or sharing medical datasets, verify that you are authorized to do so and that the necessary technical and organizational safeguards are in place. Do not rely solely on automated de-identification or redaction functions to ensure compliance with data protection requirements.

Liability notice

The information above explains the intended use of the software and the limitations of its functions and AI outputs. It does not exclude liability where such an exclusion is prohibited by law.

Liability is governed by the applicable statutory provisions and the validly agreed license terms. Mandatory statutory liability and non-waivable statutory rights remain unaffected.

Dataset citation: National Lung Screening Trial Research Team. (2013). Data from the National Lung Screening Trial (NLST) [Data set]. The Cancer Imaging Archive. https://doi.org/10.7937/TCIA.HMQ8-J677 License: Creative Commons Attribution 4.0 International (CC BY 4.0).