A Reproducible Protocol for Artificial Intelligence-Assisted Enhancement of Digitized Historical Neuroanatomical Illustrations

Summary

This methods article presents a reproducible, quality-controlled workflow for artificial intelligence-assisted enhancement of digitized historical neuroanatomical illustrations. The protocol defines source selection, extraction, file standardization, local super-resolution processing, paired original-enhanced archiving, technical quality-control review, and documentation of both acceptable and suboptimal outputs.

Abstract

Digitized historical neuroanatomical illustrations are valuable resources for anatomical education, historical scholarship, and digital archiving. However, their practical use in teaching material and scholarly presentation is often limited by low apparent resolution, fading, paper discoloration, compression artifacts, and reduced legibility of fine anatomical line work or labels. Artificial intelligence-assisted super-resolution may improve the visual usability of these digitized reproductions, but uncontrolled enhancement may also introduce artificial texture, altered line continuity, distorted labels, or visually plausible details not supported by the source image. Therefore, standardized and transparent procedures are required when applying artificial intelligence-assisted enhancement to historical scientific illustrations.

This article describes a step-by-step workflow for the controlled enhancement of digitized historical neuroanatomical illustrations. The protocol includes source-image selection, documentation of provenance and reuse status, image extraction, conservative preprocessing, local super-resolution enhancement using fixed settings, paired original-enhanced archiving, region-of-interest-based inspection, and structured technical quality-control documentation. The workflow is intended to improve visual accessibility for teaching, presentation, and preliminary visual inspection while maintaining a direct interpretive link between every enhanced output and its original digitized source.

The method should be regarded as auxiliary enhancement of scanned digital reproductions, not restoration of original artwork, authentication of historical sources, or recovery of lost anatomical information. Original and enhanced versions must be archived together, and enhanced outputs should be reviewed for possible artifacts before use. This protocol provides a practical, low-cost, and reproducible approach for preparing historical neuroanatomical images as paired companion visualizations while preserving caution regarding anatomical and historical integrity.

Introduction

Historical neuroanatomical illustrations have played an important role in the visual documentation, teaching, and interpretation of the nervous system. Before the development of modern neuroimaging, anatomical plates served as the primary medium for communicating observations of the brain, cerebral ventricles, cranial nerves, vascular structures, and brainstem anatomy^1,2,3. Many of these works are now accessible through digitized books, scanned archival collections, and online repositories. However, the direct use of digitized historical illustrations in modern teaching, presentations, and digital archiving is often limited by low resolution, fading, paper discoloration, compression artifacts, scan-related blur, and reduced legibility of fine line work or labels^4,5,6,7,8.

AI-assisted image enhancement and super-resolution methods offer a practical opportunity to improve the visual usability of degraded or low-resolution digitized images. Deep learning-based super-resolution algorithms can increase apparent resolution, improve edge definition, and reduce scan-related visual degradation in digital images^9,10,11,12. For historical neuroanatomical illustrations, such methods may help clarify fine anatomical lines, labels, and structural boundaries that are difficult to inspect in the original digitized reproduction. Nevertheless, historical anatomical images require caution. These illustrations are not only visual teaching materials but also scientific and historical documents. Enhancement may improve clarity but can also alter line texture, exaggerate edges, modify shading, or create artificial details if applied without a standardized workflow.

For this reason, artificial intelligence-assisted enhancement of historical anatomical illustrations should be performed as a controlled technical process rather than as unrestricted image manipulation. A reproducible protocol should define how source images are selected, how provenance and reuse status are documented, how images are extracted and standardized, which enhancement settings are applied, how output files are archived, and how enhanced images are checked for possible artifacts. Original and enhanced versions should remain linked throughout the workflow so that the enhanced output can always be interpreted as a processed digital reproduction rather than as a restored or historically original image.

A standardized workflow is preferable to ad hoc artificial intelligence-assisted enhancement because it defines source selection, conservative preprocessing, fixed enhancement settings, paired original-enhanced archiving, and post-enhancement quality control before the processed image is used. The intended end users include anatomists, neurosurgeons, medical educators, historians of anatomy, medical illustrators, and digital-archive users who require visually legible companion copies while preserving direct access to the original source.

This methods article presents a step-by-step protocol for AI-assisted enhancement of digitized historical neuroanatomical illustrations. The core image-processing sequence, including image preparation, fixed 16× super-resolution enhancement, and PNG export, is summarized in Figure 1. Subsequent protocol steps include paired original-enhanced archiving, post-enhancement quality control, region-of-interest-based visual inspection, and documentation of successful and suboptimal outcomes. The protocol is not intended to claim restoration of original artwork or recovery of lost anatomical information; rather, it is intended to improve the visual accessibility of digitized reproductions while preserving the interpretive relationship between the enhanced image and the original source.

Protocol

This protocol was applied only to digitized historical neuroanatomical illustrations and did not involve human participants, patient data, animal subjects, biological samples, or identifiable personal information; therefore, institutional human-subjects research approval was not applicable. All images were obtained from public-domain, permission-cleared, or appropriately licensed historical sources.

1. Source-use considerations

1. Verify the provenance of each image before processing, including the historical author, source title, approximate publication year, and digital source.
2. Interpret enhanced outputs as processed digital reproductions rather than as restored original artwork or evidence that lost anatomical information has been recovered.
3. Archive original and enhanced files together so that the processed image can always be compared directly with the source image.

2. Selection of digitized historical neuroanatomical illustrations

1. Select digitized historical illustrations published from the sixteenth through the nineteenth centuries.
2. Include illustrations depicting neuroanatomical structures such as the brain, cerebral hemispheres, gyri and sulci, ventricular system, brainstem, cranial nerves, cerebral vasculature, or related nervous system anatomy.
3. Include only images obtained from sources with identifiable provenance and public-domain or reuse-compatible status. Confirm that each selected illustration was published between the sixteenth and nineteenth centuries and that its provenance and reuse status were individually verified before inclusion.
4. Exclude duplicated images, severely cropped images, incomplete plates, modern redrawn reproductions, or images with insufficient source clarity, including source files with a long-axis dimension below approximately 1,000 pixels when a higher-quality version is available, missing or cropped anatomical content affecting more than approximately 20% of the illustration, unreadable labels in the selected region, severe compression artifacts, or scan-related degradation that prevents comparison with the original image.
5. Identify duplicate illustrations by comparing the historical author, source title, publication year, plate or page number when available, repository metadata, and visual content. When duplicate scans are available, retain the version with higher resolution, more complete plate boundaries, and clearer labels.
6. Assign each included image a unique file identity before processing.
7. Maintain paired original and enhanced versions of each image throughout the workflow.

3. Image extraction from digitized sources

1. Open the digitized source file using the highest-quality version available.
2. Retrieve each target illustration by exporting it from the source portable document format (PDF) file or by capturing it from the digital source when direct export is not feasible.
3. Use a PDF reader or export software for PDF-based image extraction. Use a standard screen capture utility only when direct export is not feasible.
4. For PDF-based extraction, use the original embedded image resolution when available and avoid downsampling during export. For screenshot-based extraction, display the source at 100% or 200% magnification and record the display resolution, display magnification, monitor scaling setting, and resulting screenshot dimensions.
5. Save the extracted image in Joint Photographic Experts Group (JPEG) or Portable Network Graphics (PNG) format, depending on the source availability and extraction method.
6. Preserve the full anatomical illustration during extraction whenever possible. If the full illustration cannot be preserved, apply the exclusion criteria in step 2.4 and document the crop.
7. Retain labels and historically relevant markings if they are part of the main illustration. Do not manually redraw, recolor, reconstruct, or retouch anatomical structures during extraction.

4. Preprocessing and file standardization before enhancement

1. Review each extracted image before enhancement.
2. Crop only non-informative outer margins or blank surrounding areas. Do not perform manual visual correction other than cropping. Specifically, do not apply manual sharpening, denoising, recoloring, redrawing, selective cleaning, or anatomical retouching before enhancement.
3. Use image editing software only for conservative image cropping and simple layout preparation. Do not perform anatomical retouching.
4. Preserve baseline images as the original reference files before enhancement. Do not force baseline image size and resolution into a single uniform pixel dimension before processing because original source dimensions vary across historical plates.
5. Confirm that each preprocessed file remains a faithful digitized reproduction of the original source image before proceeding to AI-assisted enhancement.

5. Artificial intelligence-assisted enhancement

1. Open the standardized input image in locally installed artificial intelligence-based image upscaling software.
2. Use the single-image processing mode unless batch processing is specifically required.
3. Select the input image using the Select Image function.
4. Select a fixed super-resolution model or preset before processing and use the same model for all included images.
5. Set the image scale to 16x to reproduce the fixed setting used in this workflow, and record that default software parameters were not manually modified. Do not interpret the selected scale as optimized or superior to alternative scale factors.
6. Use the same enhancement settings for all included images to maintain consistency across the workflow.
7. Set the output folder to the same directory as the input image or to a predefined project output folder, and use a separate output directory when possible to prevent accidental overwriting of original files.
8. Perform the enhancement on a computer equipped with a 13th-generation 16-core, 24-thread mobile central processing unit, a dedicated laptop graphics processing unit with 8 GB of graphics memory, and 32 GB of RAM. Enable graphics processing unit (GPU) acceleration during image processing to reduce processing time to approximately 3 min per image, depending on input image dimensions.
   NOTE: Internet access will not be required after the software and model files have been installed locally.
9. Run the enhancement process using the software’s enhancement or start-processing command.
10. Export the enhanced output in PNG format.
11. Preserve the automated output filename generated by the software when possible, or assign a structured filename that records the source image identity, enhancement scale, and output format.
    NOTE: Do not manually redraw, retouch, recolor, or selectively correct anatomical details after AI enhancement.
12. Archive the enhanced PNG file together with the original baseline image.

6. Post-enhancement quality-control review

1. Open each enhanced image together with its corresponding original image. Confirm that the enhanced file was generated completely and that no file corruption or incomplete export occurred.
2. Compare the enhanced image with the original image at low and high magnification. Check whether the enhancement appears to improve the visibility of line work, anatomical contours, labels, and scanned details.
3. Apply a structured technical quality-control checklist. This checklist is a documentation step rather than a formal expert-rater validation study.
4. Specifically inspect the output for the following enhancement-related problems: artificial texture formation; excessive sharpening, including thickened fine lines, exaggerated paper grain, or edges that appear more pronounced than in the original image; distortion of letters or labels; false continuity between separate lines; thickening of fine anatomical lines; exaggerated paper texture; and structures that are not clearly supported by the original image.
5. Classify each output as technically acceptable or suboptimal.
   1. Consider an output technically acceptable if anatomical contours, labels, and historical markings remain visually consistent with the original digitized image.
   2. Consider an output suboptimal if enhancement produces misleading texture, distorted text, artificial line formation, or excessive alteration of the historical appearance. Retain suboptimal outputs for documentation when they illustrate limitations of the method.
6. Do not replace the original image with the enhanced image. Use the enhanced image only as a processed companion version.

7. Region-of-interest selection for paired comparison

1. Select one or more representative regions of interest (ROIs) manually from each image pair after the original and enhanced full images have been archived. Choose regions that demonstrate the practical effect of enhancement on neuroanatomical visualization. Prefer regions containing visually complex or depth-rich anatomical details, such as ventricular structures, gyri and sulci, vascular line work, brainstem contours, cranial nerve-related details, or small labels. Avoid selecting only the visually best regions; include regions that fairly represent the strengths and limitations of the enhancement.
2. For each selected region, crop the same anatomical area from the original and enhanced images using identical proportional coordinates. If exact coordinate matching is not possible due to differing image dimensions, document the adjustment. Use an ROI size of approximately 10%–20% of the image width and height, adjusted only when necessary to include the same anatomical structure or label in both panels.
3. Ensure that the original and enhanced ROI panels show the same anatomical field. Do not apply additional sharpening, contrast adjustment, or manual correction to the ROI panels after cropping.
4. Use the ROI panels only for visual demonstration of the enhancement effect, not as independent expert-rating material.

8. Preparation of representative figures

1. Prepare one workflow figure summarizing the complete protocol.
2. Prepare representative paired image figures showing original and enhanced versions.
3. For each representative image figure, use a four-panel structure when possible. Place the original digitized illustration in panel A; the AI-enhanced illustration in panel B; the magnified ROI from the original image in panel C; and the matched magnified ROI from the enhanced image in panel D.
4. Use consistent panel spacing, alignment, panel labels, and magnification across paired original-enhanced panels. Use a standard sans-serif font at a consistent, legible size for panel labels and annotations, and record the selected font and size.
5. Use arrows or arrowheads only when necessary to indicate a specific line, label, or artifact, and do not add arrows that imply anatomical interpretation not visible in the original image.
6. Add scale bars only when the source image provides reliable spatial calibration. For historical plates without a valid spatial scale, do not add scale bars.
7. Do not apply any additional brightness or contrast adjustments when preparing the paired original and enhanced panels. Export figures at 300 dpi, or 600 dpi when fine-line work and labels require a higher-resolution display.
8. Include examples from different historical sources to demonstrate that the workflow can be applied across variable drawing styles and scan qualities. Include at least one example showing a suboptimal enhancement outcome, such as artificial texture formation or text distortion, to demonstrate the limitations of the method.
9. Export figures as separate high-quality PDF, TIFF, or JPEG files without embedded legends. For PDF export, preserve the source image dimensions and avoid downsampling or lossy compression.
10. Provide figure legends only in the manuscript file.
11. Check all figure panels for cropping, resolution, and panel completeness before submission.

9. Documentation and data management

1. Store the original and enhanced versions of each image as paired files so that every enhanced output can be traced back to its corresponding source image.
2. Record the source repository for each illustration, including the archive or repository name, stable URL or accession information when available, reuse status, and access date.
3. Record the processing method for each image, including the extraction method, preprocessing step, enhancement software, software version, selected model, scale factor, output format, and final file name.
4. Record whether each image was extracted by PDF export or screenshot capture, whether cropping was limited to non-informative margins, and whether enhancement used the fixed 16x super-resolution setting.
5. Preserve the software-generated output file names when possible, or assign structured file names that record the source image identity, enhancement scale, and output format.
6. Document whether each enhanced output was technically acceptable or suboptimal. Suboptimal findings may include artificial texture formation, excessive sharpening, distorted labels, or visually misleading line changes.

Results

Representative original and enhanced image comparisons are provided in the manuscript figures. The publication date, source provenance, repository information, and reuse status were verified and documented for each of the 40 historical illustrations included in this study.

Image dataset and source characteristics
The protocol was applied to 40 digitized historical neuroanatomical illustrations. The publication date, provenance, and reuse status of every illustration were i...

Discussion

This methods article presents a reproducible workflow for AI-assisted enhancement of digitized historical neuroanatomical illustrations. Historical anatomical images have long served as important tools for anatomical documentation, teaching, and scientific communication, and many of these works are now accessible through digitized books, scanned archives, and online repositories. However, their use in modern teaching, presentations, and digital archiving is often limited by low apparent resolution, faded line work, backg...

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
Adobe Acrobat	Adobe	Version not recorded	PDF-based image extraction
Computer workstation	Not reported in submitted files	Configuration not reported	Hardware used for local AI-assisted image enhancement
Display/screenshot settings	Windows screen capture tool or equivalent	100% or 200% display magnification; monitor scaling recorded	Standardized screenshot-based extraction when direct PDF export was not feasible
Gallica	Bibliothèque nationale de France	Online repository	Source of digitized historical works
General Photo Real-ESRGAN preset	Upscayl / Real-ESRGAN-based model	16x setting	Super-resolution enhancement of digitized illustrations
Google Books	Google	Online repository	Source of digitized historical books
Graphics processing unit	Not reported in submitted files	Model and acceleration status not reported	GPU acceleration for image upscaling when enabled
Internet Archive	Internet Archive	Online repository	Source of public-domain digitized historical works
Local storage drive	User-specified workstation	Sufficient free storage space	Storage of original, enhanced, and figure-ready image files
PDF export settings	Adobe Acrobat or equivalent PDF reader/export software	Original embedded resolution when available	Extraction of images from digitized PDF sources without intentional downsampling
PhotoScape X	MOOII Tech	v4.1.1	Conservative cropping and figure preparation
PNG image format	—	—	Output format for enhanced images
Processing log spreadsheet	Microsoft Excel or equivalent spreadsheet software	Version not critical	Documentation of source, extraction method, enhancement setting, processing time, output file name, and QC classification
Quality-control checklist	Author-generated checklist	Not applicable	Structured technical review for artificial texture, excessive sharpening, label distortion, line thickening, and unsupported visual changes
System memory	Not reported in submitted files	RAM capacity not reported	Memory available for high-scale image enhancement
Upscayl	Upscayl	v2.11.5	Artificial intelligence-assisted image upscaling
Wellcome Collection	Wellcome Collection	Online repository	Source of public-domain digitized historical works
Wikimedia Commons	Wikimedia Foundation	Online repository	Source of public-domain or reuse-compatible images
Windows operating system	Microsoft	Windows 10	Operating system used for image extraction and processing
Windows screen capture tool	Microsoft	Windows 10 built-in tool	Screenshot-based image capture when PDF export was not feasible