Comparison and image analysis of different file formats on whole-slide images for colorectal and non-small cell lung carcinomas

Abstract

Different scanners employ various image acquisition techniques and may store scanned images in proprietary formats or export them to a range of standard file formats. Understanding these technical limitations is essential for ensuring optimal scan quality and accurate image analysis. The present study compared image formats from two whole-slide imaging platforms—Leica's Aperio AT2 (.SVS) and 3DHistech's Panoramic P250 Flash III (.MRXS)—using artificial intelligence (AI) algorithms to quantify pixel-level, color-profile, and segmentation differences on 10 colorectal carcinoma (CRC) and 10 non-small cell lung carcinoma (NSCLC) cases. The same slides were scanned independently on both platforms at 20×, and native MRXS files were also converted to SVS using SlideMaster®, allowing us to isolate the effect of file-format translation (native MRXS vs. MRXS-to-SVS, same scanner) from that of scanner hardware (3DHistech-derived formats vs. native Aperio SVS). Native MRXS and MRXS-to-SVS files showed very high similarity in pixel values, color profile, and AI predictions, whereas native Aperio SVS files differed significantly across image size, pixel size, pixel number, and color profile—reflecting the combined effect of scanner hardware and file format rather than format alone. Given the limited sample size and absence of a pre-specified power analysis, these findings are exploratory and hypothesis-generating. We anticipate that an enterprise-level communication standard will emerge to harmonize image formats and support consistent, reproducible analysis by pathologists and AI software alike.

Introduction

The field of pathology is undergoing a paradigm shift with the integration of technologies such as digital imaging, image analysis, and computational techniques. Digital pathology (DP) is a multi-step process that depends on the smooth and seamless functioning of scanners, an image management system, integration with Laboratory Information system (LIS)/artificial intelligence (AI) systems, calibrated monitors, organizational IT support, and the training of personnel, including pathologists. Scanning of glass slides is the genesis of DP, converting analog glass slides mounted with stained paraffin tissue sections into a digital image format that can be shared, scrutinized by pathologists or image analysis software, and stored.¹ Whole-slide imaging (WSI) systems can increase workflow efficiency in pathology by enabling real-time or asynchronous collaboration, efficient access to archived slides, streamlined case handling, and access to image analysis and computational tools.²

The overall WSI process, including tissue preparation and staining as well as digitization and analysis, can introduce variation at each stage. Different scanners use various methods of image acquisition and often save images in proprietary file formats or convert them to other formats.2, 3 Consequently, the same tissue sample may appear subtly different when viewed on different scanners or viewing software, which can lead to sub-optimal evaluation by a pathologist or sub-optimal performance of an algorithm or dataset.3, 4 Variation among WSI scanners is driven by hardware differences—such as objective lenses, cameras, motor control, and light sources—as well as software differences in data handling, including image alignment, gamma correction, and file storage techniques for compression and sampling at various magnifications.⁵ A clear understanding of these technical limitations is critical to ensuring high-quality scanning and analysis. To support informed decisions and judicious investment in DP technology, we evaluated the comparative technical performance of the image formats created by the Leica Aperio AT2 and 3DHistech scanners at IQVIA labs, using AI-based image analysis on WSIs of colorectal carcinoma (CRC) and non-small cell lung carcinoma (NSCLC).

To rigorously evaluate these DP workflows, it is essential to isolate variations introduced by software-level file architecture from those driven by physical scanner hardware. Therefore, a key component of our study design involved utilizing the 3DHistech software converter, SlideMaster, to translate native MRXS files into the SVS format. By maintaining a constant physical scanner origin (the 3DHistech platform) while altering only the file wrapper format, we established a controlled baseline. This strategy allows us to evaluate whether file translation protocols alter pixel features or color integrity independently of physical optical configurations, hardware sensors, or illumination paths.

The primary endpoint of this study was to quantify pixel-level, color-profile, and AI-derived segmentation differences between native MRXS files and MRXS-to-SVS converted files, both originating from the same 3DHistech scanner, thereby isolating the effect of software/file-format translation from physical scanner hardware. A secondary endpoint was to characterize the corresponding differences between these 3DHistech-derived formats and native SVS files acquired on the Leica Aperio AT2 scanner, reflecting the combined effect of scanner hardware and file format. Each figure in the Results section is organized to address one of these two comparisons directly.

Pixel fidelity is important in DP, as it ensures the clarity and detail needed for pathologists to distinguish stain colors and tissue architecture when diagnosing disease.⁶ We therefore also evaluated whether pixel fidelity was maintained across image file format conversions between Leica's SVS and 3DHistech's MRXS formats. High pixel fidelity is essential for both human evaluation and AI analysis.

Previous studies have reported scanner-associated variability arising from differences in optics, illumination systems, color calibration, and image-processing pipelines. However, direct comparisons of native MRXS files, software-converted MRXS-to-SVS files, and native Aperio SVS files using AI-assisted, pixel-level quantitative metrics remain limited. In this study, the parameters assessed included pixel size, image size/pixel number, RGB intensity values, RGB color ratios, color profile differences, and AI-derived segmentation outputs for tissue, tumor, and necrosis compartments.

Materials and methods

The transition from glass to digital workflows requires conventional stained paraffin tissue sections on glass slides, DP equipment, image management systems, data management software, storage capacity, and trained personnel. The present study compares the different image formats of two scanning platforms—Leica Aperio AT2 and 3DHistech P250 Flash III—using AI algorithms to highlight pixel-level differences on WSIs of CRC and NSCLC samples.

Dataset

The dataset comprised 10 CRC and 10 NSCLC cases. The tissue was sourced from Indivumed GmbH, Germany. A board-certified pathologist selected the WSIs, and all tumor grades were included. Cases of CRC and NSCLC were reviewed, and slide selection was performed to include the tumor and a rim of normal native tissue.

The slides were then independently scanned by the Aperio AT2 scanner and 3DHistech Panoramic Flash 250 III scanner at 20× magnification. Scanners were appropriately calibrated and quality controlled. Slide conversion software from 3DHistech called SlideMaster® was then used to convert the MRXS images to SVS images.

To minimize biological and lab pre-analytical confounders, the same physical hematoxylin and eosin (H&E)-stained glass slides were scanned sequentially on both scanner platforms. Consequently, tissue source, fixation, tissue processing, paraffin embedding, microtomy, staining protocol, staining batch, and slide preparation remained identical across all image acquisitions. No additional tissue sections were generated between scans. Therefore, observed differences are attributable predominantly to scanner hardware characteristics, image-acquisition workflows and file-format-associated effects rather than biological or lab variability.

Pixel size metadata for each image type and tumor cohort are summarized in Table 1. These values were recorded as part of the image-format comparison and were used to contextualize downstream pixel-level and color-profile analyses.

Table 1.

Pixel sizes spanning the different image types and subgroups.

S. No	Image type	Organ	Pixel size (μm)
1.	MRXS	CRC	0.4967
2.	MRXS	NSCLC	0.4967
3.	MRXS to SVS converted	CRC	0.4967
4.	MRXS to SVS converted	NSCLC	0.4967
5.	SVS	CRC	0.5027
6.	SVS	NSCLC	0.5027

This multi-format workflow was implemented with the specific rationale of partitioning physical hardware disparities from software-driven file profiling. Comparing the native MRXS scans directly to their software-converted SVS counterparts allows for a controlled validation of the file translation algorithm's pixel fidelity. Consequently, any pixel-level or color profile deviations detected when comparing these cohorts to the native Leica Aperio SVS scans can be confidently attributed to true physical scanner hardware variations rather than format compression or translation artifacts. Regions of interest (ROIs) were delineated manually by the pathologist. Folds, bubbles, and other artifacts were excluded from the analysis. The annotations included regions of normal tissue, regions of tumor tissue, and necrosis. Supporting annotations were done by Aiforia staff (AL) and reviewed by the pathologist (MLS).

Automated tissue segmentation and pixel extraction

To achieve a high-density, automated comparison of image features across platforms without the subjectivity of manual cell counting, deep learning models were utilized strictly as automated pixel-level extraction masks. The objective of these models was to systematically isolate distinct tissue compartments, specifically total tissue, tumor tissue, and necrosis (See Supplementary Figs. 1A and 2A), so that raw physical metrics could be objectively extracted from identical ROIs. The AI was not designed or evaluated as a clinical diagnostic tool. Segmentation masks (See Supplementary Figs. 1B and 2B) were created using Aiforia Create version 5.7.1 (Aiforia Technologies, Plc), utilizing deep convolutional neural network (CNN) architectures. For both CRC and NSCLC cohorts, a multi-layer parent–child layer tree was constructed. First, a parent tissue detector layer was trained using a complex network depth and a 60-μm field of view. This was followed by a “very complex” sub-layer depth with a 50-μm field of view to differentiate nested tumor tissue from focal areas of necrosis. Training was executed over 3000–5000 iterations (spanning approximately 1–2 h) until network convergence was achieved, defined by training loss stabilizing below 1.0. Network hyperparameters were held at platform defaults, with targeted augmentations for scale, aspect ratio, maximum shear, tonality, and contrast adjusted by 10% to maximize edge sensitivity across various scanners. The final mini-batch size per iteration was fixed at 20. Satisfactory segmentation performance was mathematically established by validating the automated masks against ground-truth boundaries annotated by a board-certified pathologist. The trained network models achieved superior spatial convergence, with precision, sensitivity, and F1 scores exceeding 90% across all classes. Specifically, the native SVS CRC model achieved a parent Tissue F1 score of 99.98% and a sub-layer F1 score of 99.62%; the native SVS NSCLC model yielded a Tissue F1 score of 99.88% and a sub-layer F1 score of 99.71%. Converted architecture models for CRC and NSCLC achieved sub-layer F1 scores of 95.51% and 99.64%, respectively. Total area error across all trained segments remained vastly below our pre-established validation threshold of 7%, ranging tightly between 0.01% and 5.0%. To evaluate cross-format transferability, the model trained on MRXS-to-SVS converted images was applied to native MRXS images without further retraining, as both image sets originated from the same 3DHistech scanner. Segmentation outputs for native MRXS images were reviewed by an experienced board-certified pathologist for concordance across tissue, tumor, and necrosis compartments before RGB intensity extraction. No separate quantitative validation against independent native MRXS annotations was performed. Once the boundaries were validated, the raw red, green, and blue intensity values of the encapsulated aggregated pixels were extracted for primary downstream comparison.

For both the layers and all the classes, Aiforia's default hyperparameters were used. Table 2 outlines all the details. Aiforia platform uses its proprietary Architecture and AI engines for the training and AI model creation. The current AI Model created using Aiforia Create is CNN-based architecture. Most of the previously published Aiforia AI models can be found using the link https://www.aiforia.com/aiforia-publications.⁷

Table 2.

Aiforia AI model training hyperparameters and augmentation settings.

Category	Parameter	Value
Training settings	Weight decay	0.0001
Training settings	Initial learning rate	1
Training settings	Maximum final learning rate	0.0001
Training settings	Iterations without progress	Default
Training settings	Mini-batch size	Default
Training settings	Mini-batches per iteration	20
Network settings	Neural network structure version	Default
Network settings	AI engine version	0
Network settings	Width scaler	1
Network settings	A priori weight	0.5
Training window	Training window size	256 px/2048 px
Augmentation settings	Scale	−1/1.01
Augmentation settings	Aspect ratio	1
Augmentation settings	Maximum shear	1
Augmentation settings	Tonality luminance	−1/1
Augmentation settings	Tonality contrast	−1/1.01
Augmentation settings	Maximum white balance change	1
Image-quality settings	White noise	0
Image-quality settings	JPG compression percentage	0.5%
Blur settings	Maximum pixels	1
Blur settings	Percentage	0.5
Rotation/flipping settings	Rotation angle	−180/180
Rotation/flipping settings	Flipping	Enabled
Post-training settings	Gain factor	1

For each image format, RGB intensity values were extracted from AI-defined tissue, tumor and necrosis compartments. The values plotted in Fig. 1, Fig. 2 represent mean RGB intensity values for each compartment and image format.

Fig. 1.

A: Shows layer tree structure made in Aiforia Create. B: Shows the mean RGB intensity values extracted from AI-defined tissue, necrosis, and tumor compartments in CRC whole-slide images across the three image formats. Overall, higher values for red color profiles were seen, and lower values for blue and green color profiles on the aggregated pixels for all three detection classes. Significant differences observed for G Necrosis: MRXS vs SVS (p = 0.0172), R Necrosis: MRXS vs SVS (p < 0.0001), B Tissue: MRXS vs SVS (p = 0.0014), G Tissue: MRXS vs SVS (p = 0.0001), MRXS to SVS vs SVS (p = 0.0002), R Tissue: MRXS vs SVS (p < 0.0001), MRXS to SVS vs SVS (p < 0.0001), B Tumor: MRXS to SVS vs. SVS (p = 0.0449), G Tumor: MRXS vs SVS (p = 0.0156), MRXS to SVS vs. SVS (p = 0.0127), R Tumor: MRXS vs. SVS (p = 0.0004), MRXS to SVS vs. SVS (p = 0.0004). Statistical comparisons between image formats were performed using one-way ANOVA followed by Tukey's multiple comparisons test. Reported p-values are Tukey-adjusted p-values. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 2.

Shows the mean RGB intensity values extracted from AI-defined tissue, necrosis and tumor compartments in NSCLC whole-slide images across the three image formats. Overall, higher values for red color profiles were seen, and lower values for blue and green color profiles on the aggregated pixels for all three detection classes. Significant differences observed for G Necrosis: MRXS vs. SVS (p = 0.0057), MRXS-to-SVS vs. SVS (p = 0.0399), MRXS to SVS vs. SVS (p = 0.0464), G Tissue: MRXS vs. SVS (p = 0.0023), MRXS-to-SVS vs. SVS (p = 0.0024), R Tissue: MRXS vs. SVS (p < 0.0001), MRXS-to-SVS vs. SVS: (p < 0.0001), B Tumor: MRXS vs. SVS (p = 0.0011), MRXS-to-SVS vs. SVS (p = 0.001), G Tumor: MRXS vs. SVS (p < 0.0001), MRXS-to-SVS vs. SVS (p < 0.0001), R Tumor: MRXS vs. SVS: (p < 0.0001), MRXS-to-SVS vs. SVS: (p < 0.0001). Statistical comparisons between image formats were performed using one-way ANOVA followed by Tukey's multiple comparisons test. Reported p-values are Tukey-adjusted p-values. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

SlideMaster conversion

Native MRXS files were converted into SVS format using SlideMaster® software using standard vendor-recommended settings. Metadata and image hierarchy were preserved during conversion. User-accessible conversion parameters were not modified during the process. The purpose of the conversion workflow was to evaluate whether file-format translation introduced measurable pixel-level or color-profile variation independently of scanner hardware.

SlideMaster does not expose user-configurable compression parameters for MRXS-to-SVS conversion; it applies a fixed, non-negotiable lossy JPEG (ISO/IEC 10918-1) codec at a quality setting of Q  =  80. This conversion is therefore not guaranteed to be pixel-lossless: native MRXS files store image data as overlapping tiles, which must be decompressed and re-tiled into the non-overlapping tile scheme used by SVS, and JPEG re-encoding is not bit-exact across generations even when the nominal quality setting is matched.¹⁶

Statistical analysis

Statistical analyses were performed separately for CRC and NSCLC. For each tumor type, quantitative image-analysis outputs were compared between the three image formats: native MRXS, MRXS-to-SVS converted images, and native SVS. The analyzed variables included RGB intensity values and derived color ratios extracted from AI-defined tissue compartments, including tissue, tumor, and necrosis.

Comparisons between image formats were performed using one-way ANOVA followed by Tukey's multiple comparisons test for post-hoc pairwise comparisons. Tukey-adjusted p-values are reported in the figure legends. A Tukey-adjusted p-value <0.05 was considered statistically significant. Results are presented as mean values, with mean differences and 95% confidence intervals reported where applicable.

Given the exploratory nature of the study and the modest sample size, p-values were interpreted cautiously and in conjunction with the direction and magnitude of observed differences.

Results

The AI-based segmentation approach enabled comparison of tissue, necrosis, and tumor detections between image formats for both the CRC and NSCLC cohorts.

Mean RGB intensity values extracted from AI-defined tissue, necrosis, and tumor compartments in CRC WSIs are shown in Fig. 1. Across all three compartments, red channel intensities on the aggregated pixels were consistently higher than blue and green channel intensities, with the most pronounced differences observed for necrosis.

A comparable pattern was observed for NSCLC: mean RGB intensity values extracted from AI-defined tissue, necrosis, and tumor compartments (Fig. 2) again showed higher red channel intensities and lower blue and green channel intensities across all three compartments, with the largest differences again observed for necrosis.

On visual inspection, a marked color-balance shift was apparent between image formats. To characterize this shift quantitatively, RGB color ratios were calculated from the aggregated pixel intensities extracted from the tumor and necrosis masks, because H&E-stained images are composed of red, green, and blue color signatures. These ratio comparisons (Fig. 3, Fig. 4) confirmed that red-channel dominance, relative to blue and green channels, was significantly higher in MRXS and MRXS-to-SVS converted images than in native SVS images, for both CRC and NSCLC. These ratio-based comparisons provide an initial quantitative framework intended to inform future functional and morphological studies of scanner- and format-associated color variation.

Fig. 3.

Ratio calculations of the different color profiles spanning the different image types (CRC).

Fig. 4.

Ratio calculations of the different color profiles spanning the different image types (NSCLC).

Representative CRC and NSCLC WSIs across the three image formats are shown in Fig. 5, Fig. 6, respectively. These images illustrate the visible color-balance differences observed between native Aperio SVS files and 3DHistech-derived MRXS/MRXS-to-SVS files.

Fig. 5.

WSI from colorectal carcinoma across the three image formats: (a): CRC: SVS (Aperio AT2) (b) MRXS (3DHistech P250 Flash III) (c) MRXS to SVS. Scale bars are shown on representative panels to avoid overcrowding the figure: main image A = 2.5 mm; inset b = 400 μm.

Fig. 6.

WSI from NSCLC across the three image formats: (a): NSCLC: SVS (Aperio AT2), (b) MRXS (3DHistech P250 Flash III) (c) MRXS to SVS. Scale bars are shown on representative panels to avoid overcrowding the figure: main image A = 2.5 mm; inset b = 400 μm.

Table 1 shows pixel sizes spanning the different image types and subgroups. Pixel size was consistent within each image type across the CRC and NSCLC cohorts. However, pixel size of SVS differed slightly as compared to MRXS and MRXS-to-SVS converted images.

MRXS and MRXS-to-SVS converted files had very high levels of similarity at a pixel level, color profile, and AI prediction, whereas native SVS files differed significantly across these measures, including image size, pixel size, pixel number, and color profiles of aggregated pixels. The outcome comparisons were performed efficiently at the whole-slide level using a CNN-based deep learning approach.

Discussion

There has been an exponential increase in the use of DP in recent years. Digital technologies are transforming modern pathology, making it efficient and ready for precision medicine. An efficient workflow is dependent on the interplay between different scanners and IT systems. Proprietary file formats, such as Aperio's. SVS and 3DHistech's. MRXS are commonly used in WSI.³ Whereas SVS has become a de facto standard due to its widespread use and openly documented format, many scanners still produce proprietary formats that may not fully adhere to FAIR (findable, accessible, interoperable, and reusable) principles.⁴ The challenge of managing and analyzing these large image files, often exceeding 20 gigabytes, is further compounded by the need for efficient storage and optimal image data compression.2, 3 AI offers transformative potential in pathology, with histopathological images remaining the diagnostic gold-standard.⁷ However, the effectiveness of AI in DP is often limited by technical inconsistencies in the production of WSIs, particularly concerning color calibration and pixel fidelity.8, 9, 10 These inconsistencies can impact pathology accuracy, AI performance, and diagnostic confidence.⁸ Until a uniform standard image format language is adopted across the industry, recognition of these differences will be paramount.

In our study, the 3DHistech software converter SlideMaster® was used to convert the MRXS images to SVS images. We found that MRXS and MRXS-to-SVS converted files had very high level of similarity on a pixel level, color profile level, and also AI prediction compared to SVS files. MRXS is a proprietary WSI file format originally developed by Carl Zeiss MicroImaging and subsequently adopted by 3DHistech scanning systems. MRXS files are known for their high image quality and support for multiple image layers and annotations. Typically, all 3DHistech scanners export in this format, and have built-in software to convert to different image formats like Digital Imaging and Communications in Medicine (DICOM), Tiled TIFF, and SVS—the object of this study. The scanner also allows for conversions to MRXS from TIFF, JPEG, PNG, DICOM etc. We observed high levels of concordance between the MRXS and the MRXS-to-SVS image formats.

Harmonization among image formats assures consistent, reproducible analysis of images, whether by a pathologist or AI software. An enterprise-level communication standard, DICOM, has been proposed, and scanner manufacturers and other organizations are actively working to move the field toward this direction.4, 9, 11 The DICOM standard now provides support for WSI digital slides, incorporating a way to handle tiled large images as multi-frame images and multiple images.¹³ This approach is particularly useful for WSI because it avoids the unnecessary transfer of large amounts of data over the network and allows efficient access to image data as well as associated metadata.9, 12 The integration of DICOM for DP is a core requirement for its success, as it provides a definition for Whole Slide Microscopy Images, often referred to as a “Supplement 145 image”.¹³ DICOM provides the ease of inter-operability, and enterprise integration, however, the adoption of DICOM has been slow due to many challenges including a large file size, and slower scanning speed among others.⁹

In addition to diagnostic evaluations, understanding and evaluating all the technical parameters of DP is essential for the successful deployment of DP. Apart from the mandatory validation of DP, day-to-day quality control and tracking of pre-imaging factors such as slide preparation, tissue contaminants, personnel training, and imaging factors like scanner and display monitor calibration are essential for the successful deployment of imaging for diagnosis, review, and development of AI algorithms.8, 9, 10 Physical color calibration of DP scanners is vital for robust AI-assisted cancer diagnosis, as technical inconsistencies can limit the reliability and applicability of AI in diverse clinical settings.¹⁰

Our findings, demonstrating significant differences between SVS files from Aperio and MRXS/MRXS-converted SVS files from 3DHistech in terms of image size, pixel size, pixel number, and color profiles, underscore the importance of these technical considerations. These variations can directly impact the performance and reliability of AI algorithms, as evidenced by the differences in AI prediction outcomes observed in our study. The high similarity between native MRXS and MRXS-converted SVS files suggests that the conversion process within the 3DHistech ecosystem maintains pixel fidelity and color integrity, which is crucial for consistent AI analysis.

Impact of scanner variability on AI

The significant differences observed in pixel size and color profiles between the Aperio AT2 and 3DHistech scanners (Table 1) highlight the challenges of AI generalizability. AI models trained on one scanner's output may not perform as well on another's due to these “domain shifts.”14, 15 For instance, the higher red color profiles in MRXS images compared to SVS images could lead to variations in how a CNN identifies tumor margins or necrotic regions.

Our results showed that significant differences were observed for necrosis and tumor detection between the two platforms. This aligns with recent research indicating that lossy conversion or hardware-specific color rendering can affect AI performance.16, 17 Specifically, the intensity values and color gradients (Fig. 1, Fig. 2) are critical for AI algorithms that rely on pixel-level features for segmentation.

Beyond their technical relevance, these format- and scanner-associated differences carry potential clinical significance. Necrosis and tumor-margin delineation, the compartments most affected by color-profile shifts in this study, are frequently used to guide tumor grading, assessment of treatment response, and quantification of biomarkers such as tumor cellularity for molecular testing. Inconsistent color rendering or pixel calibration between scanning platforms could therefore contribute to inter-observer variability among pathologists, or to inconsistent AI-assisted quantification when algorithms are deployed across sites using different scanners or file formats. This is particularly relevant in multi-site clinical trials and multi-center diagnostic networks, where WSIs are frequently generated on heterogeneous scanner hardware and exchanged between institutions. Establishing scanner- and format-specific calibration or harmonization procedures before clinical or AI deployment may therefore be an important step in mitigating diagnostic discordance.

An additional limitation of this study is its limited sample size (10 CRC and 10 NSCLC cases), for which no formal power analysis was performed before data collection. As a result, the numerous statistical comparisons reported in Fig. 1, Fig. 2, although corrected for multiplicity using Tukey's method, should be interpreted as exploratory and hypothesis-generating rather than confirmatory. Replication in larger, adequately powered cohorts will be required before these findings can be considered definitive.

A limitation of this study is that quantitative validation of AI model transferability to native MRXS images was not performed. Although MRXS and MRXS-to-SVS images originated from the same scanner and showed visual concordance following review by an experienced board-certified pathologist, future work should include independent MRXS annotations and quantitative validation metrics, such as precision, sensitivity, F1 score, Dice/IoU and area error, to confirm model performance across file formats.

A further limitation of this study is that the comparison between native SVS files and MRXS/MRXS-to-SVS files involves two different scanner platforms, Aperio AT2 and 3DHistech P250 Flash III. Therefore, differences observed between Aperio SVS and 3DHistech-derived image formats cannot be attributed to file format alone, as scanner hardware, optics, sensors, illumination, acquisition pipeline, color rendering, and compression may also contribute. The most controlled comparison in this study is between native MRXS and MRXS-to-SVS converted images, as both originated from the same 3DHistech scanner. Accordingly, conclusions regarding file-conversion fidelity are limited to the MRXS-to-SVS conversion workflow, whereas comparisons with Aperio SVS should be interpreted as scanner/platform-associated differences rather than isolated file-format effects.

Conclusions

Resource and budget constraints inherent to exploratory studies of this scale necessitated trade-offs between sample size and analytical depth. Every effort was made to extract robust, high-density quantitative features from this modest dataset, providing a resource that can inform future, larger-scale studies. This study aimed to assess changes in image creation across different WSI scanners and provide an overview of the implemented alterations. Native MRXS and MRXS-to-SVS converted files showed high concordance within the 3DHistech workflow, suggesting that the conversion process preserved the assessed pixel-level and color-profile features. In contrast, Aperio SVS files differed from 3DHistech-derived files across several parameters, including image size, pixel size, pixel number, and color profiles. However, these differences should be interpreted as scanner/platform-associated effects rather than file-format effects alone, as scanner hardware, optics, illumination, acquisition pipeline, and file format may all contribute. Variabilities noted will impact visual analysis and machine learning utility, potentially hindering AI advancement and generalizability. Outlining these changes is essential for AI developers, along with establishing quality measures and testing tools following best scanning practices. Together, these findings highlight the need to evaluate scanner platform, acquisition pipeline, file-conversion workflow, and color/pixel-level consistency before deploying AI-based image analysis across different DP systems.

Funding

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Declaration of AI-assisted technologies in the writing process

During the preparation of this work, the authors used assisted technology to improve language and readability. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this article.

Appendix A. Supplementary data

Supplementary material

Representative organs with corresponding AI segmentation mask.

Data availability

The data supporting the findings of this study are not publicly available due to sensitivity concerns. Interested researchers may request access to the data from the corresponding authors, subject to reasonable justification and appropriate safeguards.