Redactwizard DICOM Research with AI
Product White Paper
A portable, locally operated tool for viewing, redacting, technically preparing, and AI-assisted exploration of medical imaging data
1. Executive Summary
Redactwizard DICOM Research is a self-contained 2.2 GB portable Windows desktop application for the local viewing, examination, and technical processing of medical imaging data. It combines a DICOM viewer with tools for selective DICOM metadata de-identification, OCR-assisted text recognition, manual and OCR-assisted pixel masking, interactive image annotation, and locally executed multimodal AI analysis.
The application is designed for research, education, training, software evaluation, and other non-clinical use cases. It supports individual DICOM files, multi-frame DICOM objects, DICOM series stored in directories, and the direct import of DICOM studies from ZIP archives. Common image formats such as JPEG and PNG can also be opened.
Redactwizard DICOM Research combines rule-based processing of supported DICOM metadata with heuristic text recognition and probabilistic AI inference. Users can view medical images, modify identifying information, mark relevant image regions, and explore visual content without relying on a cloud-based analysis service for the integrated OCR and AI functionality.
Processing takes place locally on the Windows PC. The portable desktop architecture brings image viewing, metadata de-identification, text recognition, pixel masking, AI-assisted analysis, and PDF generation together in a single working environment.
Redactwizard DICOM Research is not intended for clinical diagnosis or therapeutic decision-making. The application does not replace medical professionals, clinically validated software, or established medical reporting procedures.
2. Objectives and Scope
Medical imaging datasets often contain more information than the image pixels themselves. In addition to image data, DICOM files may contain patient identifiers, accession numbers, dates and times, institutional information, free-text comments, and manufacturer-specific attributes. Personally identifiable information may also be embedded directly in the image pixels, for example as overlaid patient identifiers, dates, or equipment labels.
At the same time, there is a growing need to view and prepare medical imaging data locally for scientific investigations, educational materials, software testing, and the evaluation of AI-assisted image processing.
Redactwizard DICOM Research addresses these needs through an integrated desktop environment that combines the following functional areas:
- Image viewing: Display of individual DICOM images, multi-frame DICOM objects, and supported image series.
- ZIP import: Direct opening of ZIP archives containing DICOM files, with local extraction to a temporary cache.
- Metadata de-identification: Selective removal or replacement of supported personally identifiable DICOM attributes.
- Text recognition: Local OCR-based recognition of text embedded directly in image pixels.
- Pixel masking: Manual and OCR-assisted masking of selected image regions.
- Image annotation and measurement: Tools for distance measurements, rectangular image regions, arrows, and text notes.
- Multimodal AI: Exploratory image descriptions and natural-language question answering about visual image content.
- Documentation: PDF export of selected images and AI-generated analysis results.
The individual functions can be combined within a local workflow. The principal editing and selection steps remain under the user's control.
3. Image Viewing and DICOM Viewer
3.1 Supported Image Data
Redactwizard DICOM Research supports the display of the following file types and image structures:
- Individual DICOM files
- Multi-frame DICOM objects
- DICOM series stored in directories
- DICOM files contained in supported ZIP archives
- PNG images
- JPEG images
The viewer supports opening multiple examinations in separate tabs. This allows different imaging datasets to be examined within a shared working environment without managing them in separate application windows.
The content that can actually be displayed depends on the file format, the structure of the DICOM object, and the implemented decoding capabilities.
3.2 Direct Import of DICOM Studies from ZIP Archives
Redactwizard DICOM Research supports the direct import of ZIP archives containing DICOM files. This is particularly useful for studies downloaded from PACS web portals or research repositories such as The Cancer Imaging Archive (TCIA).
Manual extraction through Windows File Explorer is not required for this import workflow. Instead, users can select a ZIP file directly through the application's file dialog. Redactwizard identifies the archive by its file extension and routes it to the local ZIP-processing functionality.
The archive is then extracted into a temporary cache directory. During extraction, archive paths that would escape the designated destination directory are rejected by the ZIP importer's path-validation mechanism. Unnecessary macOS metadata entries can be skipped. The extracted files are then passed to the DICOM series loader.
DICOM files that can be processed are loaded into the viewer as an image series. Users therefore do not need to organize the extracted files manually in another directory before beginning their examination.
The temporary cache serves an important technical purpose: the extracted files remain available while users navigate through the image series. The viewer can load additional frames from disk when needed instead of keeping all image data in memory simultaneously.
This approach is particularly useful for larger DICOM studies in which individual image slices are viewed sequentially. Whether every file in a particular archive can be processed depends on the archive structure and the compatibility of its contents with the implemented DICOM series loader.
ZIP import thus complements the opening of individual files and existing DICOM series with a simplified import workflow. The archive is processed locally and does not need to be transferred to an external service for import.
3.3 Image Display and Navigation
Various display and navigation functions are available for interactive examination of medical imaging data:
- Window and Level adjustments for changing brightness and contrast
- Direct display adjustments using mouse movements
- Zooming and panning of the image viewport
- Rotation in 90-degree increments
- Horizontal flipping
- Fitting the image to the application window
- Display at the original image size
- Frame navigation using buttons, the mouse wheel, and a slider
A status indicator displays the current position within an image sequence. For multi-frame objects and image series, users can move between individual frames and track their current position in the sequence.
These viewing options facilitate both general image inspection and focused examination of selected regions. The information available depends on the loaded image data and the supported DICOM attributes.
3.4 Performance and Responsiveness
Image loading is performed asynchronously in the background. An LRU-based frame cache keeps recently used image data available. Additional prefetching can prepare subsequent frames and support smoother navigation.
Time-consuming tasks are delegated to background operations or worker threads. The user interface therefore does not have to wait for image loading, OCR, or AI inference to finish before responding to further user input.
Actual performance depends on factors including image size, frame count, available memory, and hardware.
4. DICOM Metadata and De-identification
4.1 Selective De-identification
DICOM files can contain structured information that may identify a person, a medical institution, or an examination. Redactwizard DICOM Research provides a dialog-based interface for selecting supported metadata fields to be removed or modified.
The information categories covered include, in particular:
- Patient names and patient identifiers
- Accession numbers
- Dates of birth
- Addresses and telephone numbers
- Information about institutions and treating or referring personnel
- Information about examiners and operators
- Patient, study, and image comments
- Supported private and manufacturer-specific tags
Metadata de-identification can therefore extend beyond individual patient attributes to include additional structured or text-based information, provided it is stored in supported DICOM fields.
The selection is made through the user interface. This allows users to specify the attribute groups to be processed before starting the operation and to maintain control over the individual processing steps.
4.2 Identifier Consistency
DICOM objects use identifiers to link patients, studies, series, and individual instances. When these identifiers are modified, care must be taken not to unintentionally break relationships within a dataset.
Redactwizard DICOM Research supports consistent, cryptographically derived modification of selected identifiers. The aim is to preserve the intended relationships within a processed dataset as far as the implemented procedure allows, without retaining the original identifiers unchanged.
Which relationships remain intact depends on the attributes being processed and the structure of the particular dataset. This does not constitute a general guarantee of compatibility with every DICOM object type or every possible relationship between files.
4.3 Private Tags and Comments
Manufacturer-specific private tags may contain additional information that is not represented by standard DICOM fields. The software therefore also supports the de-identification of selected private tags.
Supported free-text fields are processed as well, including patient, study, and image comments. Since their contents and storage locations can vary between DICOM producers and application systems, additional checks for unsupported attributes remain necessary.
4.4 Processing Entire Series
Metadata de-identification can be applied to individual files or entire directories containing DICOM files. During series processing, an error log helps users identify problems encountered while processing individual files.
The software can also inspect the DICOM header for the BurnedInAnnotation (0028,0301) attribute and display a corresponding warning. This attribute may provide an indication of whether visible information is embedded in the image pixels. Its value does not replace direct inspection of the actual image content.
5. OCR and Masking of Embedded Information
5.1 Background
Not all personally identifiable information in medical imaging data is stored as structured DICOM metadata. Some information is part of the pixels themselves, for example overlaid patient identifiers, dates, or equipment labels.
Such information is not removed by header de-identification alone. Redactwizard DICOM Research therefore complements the processing of structured metadata with local OCR-based text recognition and manual pixel masking.
5.2 Local Text Recognition
The integrated OCR component searches for text-like image regions and operates in the background while the user interface remains available for interaction. Detected regions are highlighted directly in the viewer.
The recognition process is supplemented by geometric plausibility filters. For example, area and aspect ratio can be used to exclude obviously unsuitable candidates and reduce unwanted markings on imaging structures.
These filters are heuristic. They can reduce false detections but cannot reliably exclude every irrelevant region or guarantee that all relevant text will be detected.
5.3 Manual Pixel Masking
In addition to automatic recognition, users can draw rectangles over selected image regions. OCR-detected regions and manually defined masking areas are visually distinguished in the viewer.
Detected regions can be removed individually from the selection before masking is applied. When needed, defined masking regions can be propagated to additional images in the same series.
This allows users to review automatic suggestions, correct them manually, and verify the selection before proceeding to the next processing step.
5.4 Modification of Image Pixels
The masking function overwrites selected pixel regions in the underlying image buffer with zero values. Masked areas are therefore not merely concealed by a colored overlay.
The modification affects the image displayed in the viewer as well as subsequent AI analysis and PDF export. The purpose is to prevent already masked regions from appearing as unchanged image content in these subsequent processing steps.
The described modification applies to the processed image data within the relevant workflow. It does not automatically mean that every original source file on disk has been modified or that all other copies, exports, and embedded content associated with a dataset have also been de-identified.
Successfully masking individual image regions is not equivalent to completely anonymizing an entire DICOM dataset.
6. Image Annotation and Measurement Tools
Redactwizard DICOM Research provides interactive tools for the technical examination and annotation of image data.
Ruler: Two points can be connected to measure a distance. The interface displays the measurement. Depending on the available image information and implemented calibration, measurements may be presented in pixels or millimeters.
Arrow: A vector arrow can point to a specific image region, highlighting an area for discussion, demonstration, or subsequent documentation.
Rectangular ROI: A rectangular region of interest can be marked, with the interface displaying the dimensions of the selected region.
Text note: Text can be placed at a freely selected position in the image, for example to explain a demonstration or prepare educational materials.
The active tool is highlighted visually. Adapted cursors and a crosshair support precise positioning while drawing.
These tools are intended for technical image exploration and documentation. Measurements and annotations should not be interpreted as validated clinical measurements or medical findings. In particular, users must verify that physical scaling information is available and suitable for the dataset in question.
7. Local Multimodal AI Analysis
7.1 Functionality
Redactwizard DICOM Research integrates a locally executed multimodal Vision-Language Model for the exploratory analysis of visual image content.
The AI component can describe the currently provided image in natural language and respond to questions entered by the user. Predefined analysis templates support recurring tasks, while custom prompts allow users to formulate their own questions and instructions.
Available approaches include, among others:
- General image descriptions
- Structured and systematic image exploration
- Descriptions of potential abnormalities
- Objectively worded image summaries
- Freely formulated questions and instructions
The software can display the generation progress continuously. This allows users to follow the analysis process without having to wait for the complete output before receiving an update.
7.2 Probabilistic Results
The AI processes visual information using statistical patterns learned during model training. Its responses are probabilistic model outputs rather than deterministic assessments of every image feature.
The results may be factually incorrect, incomplete, inconsistent, or misleading. The model may overlook relevant image details, describe features that are not present, or fail to answer questions reliably.
A convincing response does not establish that the analysis is factually correct. Users must critically review the results and must not interpret them as medical findings or as confirmation of a particular disease.
7.3 Configuration
Depending on the available hardware and intended workflow, technical inference parameters can be adjusted. These include GPU layers, batch size, context length, and token limits.
These options allow local execution to be adapted to different computer configurations and tasks. Performance may vary depending on the processor, memory, available GPU acceleration, image size, and prompt complexity.
The application continuously displays a notice identifying the medical AI functionality as intended for non-clinical use.
8. PDF Export and Documentation
Redactwizard DICOM Research can combine the results of an analysis workflow into an A4 PDF document.
An exported report may include:
- The rendered image, including applied masking
- The associated AI-generated analysis
- A notice identifying the non-clinical intended use
- Selected metadata
PDF generation is based on a native PDF component and supports the creation of structured document elements. Images are incorporated according to the intended report layout.
Analysis results can also be copied directly to the clipboard for use in other applications.
Since a PDF can combine different types of content into a persistent document, users should review the completed report before sharing it. This is particularly important for embedded images, annotations, included metadata, and any personally identifiable information that may remain.
PDF export is intended for technical documentation, research, education, and software evaluation. It does not constitute the creation of a clinically validated medical report.
9. Local Architecture and Technical Implementation
Redactwizard DICOM Research is designed as a native desktop application. Its application logic is implemented in Rust.
The architecture relies on locally executable processing components and background tasks to combine image viewing, text recognition, and AI analysis within the same application.
9.1 Technology Overview
| Component area | Technical foundation | Purpose |
|---|---|---|
| Application and core logic | Rust | Native application logic and processing |
| User interface | egui | Desktop interface and interaction |
| Architecture & Package | Portable .exe | approx. 2.2 GB, incl. integrated vision model |
| DICOM processing | DICOM components and de-identification functions | Processing DICOM objects, pixel information, and supported metadata |
| ZIP import | ZIP processing with path validation | Local extraction of supported archives and loading of contained DICOM files |
| OCR and text recognition | Rust-based OCR components and tensor execution | Detection of text-like regions in image data |
| Multimodal AI | Native local inference runtime with model components | Image descriptions and natural-language question answering |
| PDF generation | Native PDF library | Creation of A4 reports |
| Background processing | Asynchronous tasks and worker threads with message passing | Background operations and a responsive user interface |
The listed components describe the technical implementation. They do not constitute a general guarantee of error-free operation, complete DICOM compatibility, or suitability for safety-critical applications.
9.2 Locally Integrated Processing
The OCR and AI components are designed to run on the local computer. No cloud-based AI service is required for these functions. The necessary model components are integrated into the intended application package.
Local execution reduces dependence on external analysis platforms and allows image data to be processed within a controlled desktop environment.
However, avoiding cloud inference does not automatically eliminate all network or system risks. Access controls, storage devices, backups, exports, the operating system, and other installed software remain part of the local security environment.
9.3 Configuration and Internationalization
The application supports German and English. Users can switch between light and dark display modes.
Settings are stored in a local configuration file that can be located alongside the executable application. This approach supports portable deployment without requiring a centralized configuration service.
10. Privacy-Oriented and Controlled Workflows
Redactwizard DICOM Research combines local data processing with user-selectable editing steps. It can support the handling of sensitive research data by bringing metadata de-identification, text recognition, and pixel masking together in one environment.
The following workflow is recommended as a general approach to controlled processing:
- Select a suitable source dataset that can be used lawfully.
- Identify the DICOM metadata fields to be processed and apply the selected de-identification steps.
- Inspect the images for embedded text and potentially identifying image regions.
- Review automatically detected OCR regions and add manual masks where necessary.
- Check the displayed image data and applied masks.
- Use AI results for exploratory purposes only and verify their content.
- Review exported reports for remaining personal or confidential information before sharing them.
This workflow is intended as practical guidance and does not replace an institutional data protection framework.
The detection and removal of personally identifiable information are subject to technical limitations. Unsupported DICOM attributes, private tags, embedded objects, undetected image text, and other data sources may continue to contain identifying information. Appropriate expert review therefore remains necessary when de-identifying real-world data.
Redactwizard DICOM Research does not guarantee complete or legally conclusive anonymization. Responsibility for the lawfulness of data processing, the review of modified datasets, and approval for their release remains with the users and the institutions responsible for the data.
11. Intended Use Cases
Redactwizard DICOM Research is designed for individuals and organizations that need to examine, prepare, or document medical imaging data for non-clinical purposes.
| Application area | Examples |
|---|---|
| Scientific research | Exploratory examination of medical imaging data and preparation of suitable research datasets |
| Education and training | Image demonstrations, teaching materials, and non-clinical exercises |
| Software evaluation | Testing image-processing workflows and exploratory evaluation of local AI functionality |
| De-identification | Cleaning supported metadata and optionally masking embedded personally identifiable information |
| Technical documentation | Annotations, image comparisons, and PDF reports for non-clinical purposes |
| Image data exploration | Viewing DICOM series, ZIP-packaged DICOM studies, and supported image files within a portable desktop application |
The use of real patient data requires compliance with applicable legal, data protection, institutional, and, where relevant, ethical requirements.
12. Distribution and Licensing
Redactwizard DICOM Research is distributed as a portable Windows application. It is designed to run without a conventional installation process.
The application integrates its intended OCR and AI components into the software package. The described local processing functions do not require a cloud-based analysis provider.
The product is offered under a perpetual license. License verification is designed to validate digitally signed license keys locally; continued operation of the local DICOM, OCR, and AI functions does not require a permanent connection to a cloud-based licensing service. The applicable usage rights are determined by the license purchased.
| Feature | Description |
|---|---|
| Product type | Portable desktop application |
| Size | approx. 2.2 GB, incl. integrated vision model |
| Target platform | Windows, 64-bit |
| User interface | Native desktop GUI |
| DICOM viewing | Individual files, multi-frame content, and series |
| ZIP import | Direct import of supported ZIP archives containing DICOM files |
| Additional image formats | JPEG and PNG |
| Metadata de-identification | Selective processing of supported attributes |
| OCR | Locally executed text recognition |
| Pixel masking | Manual and OCR-assisted |
| Image annotation | Measurement line, arrow, rectangular ROI, and text |
| AI analysis | Local multimodal image exploration |
| Report generation | PDF export and PNG Image Screenshot export |
| Cloud-based AI inference required | No |
| Languages | German and English |
| Display modes | Light and dark |
| License model | Perpetual license |
13. Limitations and Non-Clinical Intended Use
Redactwizard DICOM Research is intended exclusively for research, education, training, software evaluation, and other non-clinical applications.
The software is not intended for diagnosing, screening for, detecting, or ruling out diseases; patient monitoring; medical prognosis; treatment planning; patient triage; or decisions concerning individual medical interventions.
The locally executed AI generates probabilistic outputs. These may contain errors and must not be interpreted as expert medical assessments or clinically validated findings.
Metadata de-identification, OCR, and pixel masking are also subject to technical limitations. Information that is not detected or supported may remain in the dataset. These functions do not replace a comprehensive data protection review or independent validation of the processed data.
The application is not intended to function as a medical device for clinical purposes and must not be used to support clinical decisions outside its described non-clinical scope.
14. Redactwizard DICOM Research in Practice
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).
15. Outlook
Redactwizard DICOM Research combines DICOM image viewing, ZIP import, local text recognition, selective metadata de-identification, manual and OCR-assisted pixel masking, interactive annotation, and multimodal AI analysis in a portable Windows application.
Its native desktop architecture and locally integrated processing components enable the technical examination and preparation of supported image data without requiring cloud-based services for OCR or AI analysis.
A key strength of the application is the integration of these capabilities within a controllable local workflow. Users can load image series directly from directories or ZIP archives, modify selected metadata, review image regions containing text, create annotations, and explore AI-generated descriptions.
The software therefore supports non-clinical research, education, and evaluation tasks involving the examination, preparation, and documentation of medical imaging data.
Redactwizard DICOM Research is explicitly intended as a research and productivity tool. It does not replace established data protection procedures, medical professionals, clinically validated systems, or expert review of the generated results.
Important Disclaimer: Non-Clinical Use Only
Redactwizard DICOM Research is intended exclusively for non-clinical purposes, including research, education, training, and software evaluation. It is not intended or offered for use as a medical device, diagnostic system, clinical decision-support system, or substitute for qualified medical judgment.
The software must not be used for diagnosing diseases, screening or triaging patients, recommending or planning treatment, monitoring patients, assessing individual patient conditions for clinical purposes, or making, supporting, or influencing any clinical decision. It is not designed, validated, or intended to provide clinically reliable findings.
AI-generated outputs are experimental and probabilistic. Image descriptions, answers to questions, OCR results, and other automated outputs may be inaccurate, incomplete, misleading, or entirely incorrect. The AI may overlook relevant image features, misinterpret visual information, or generate unsupported conclusions. Outputs must not be interpreted as medical findings, medical advice, or verified clinical results.
For AI-assisted image analysis, use only synthetic, anonymized, or appropriately de-identified datasets. Functions used specifically to prepare datasets for de-identification must be limited to that purpose and used with appropriate authorization and safeguards. The software does not guarantee complete de-identification or anonymization. Users are responsible for reviewing processed datasets and checking for residual identifying information before sharing or releasing them.
Redactwizard DICOM Research does not replace qualified medical professionals, validated clinical software, established clinical workflows, or applicable regulatory and quality-assurance requirements. Human review of AI-generated outputs does not make the software suitable for clinical use.
By using the software, users are responsible for ensuring that their activities remain within its stated non-clinical purpose and comply with applicable laws, data-protection requirements, research protocols, and institutional policies.
Trademark notice: DICOM® is a registered trademark of the National Electrical Manufacturers Association (NEMA). Redactwizard DICOM Research is an independent product and is not affiliated with or endorsed by NEMA. NEMA does not sponsor or recommend the product.