Retrospective Identification and Quantitative Image Analysis of the ICAP Nuclear Pattern AC‐30 in Antinuclear Antibody Testing Authors: Changmeng Wu, Wenhan Du, Weidong Chen, Juan He, Linlin Xu, Yu Li Published: 2026 ABSTRACT ABSTRACTBackgroundThe International Consensus on ANA Patterns (ICAP) recently codified the AC‐30 nuclear pattern, characterized by fine speckled nuclear texture and metaphase chromatin staining. Although formally recognized, its clinical significance remains unclear. Quantitative methods and objective outcome measures for characterizing AC‐30 have not been established.MethodsA retrospective analysis was conducted using archived HEp‐2 ANA images. AC‐30 was defined by unanimous scoring from three blinded experts. AC‐1 and AC‐2 served as comparator sets. Clinical diagnoses and serological data were retrieved from hospital records. Pixel classification was performed using ilastik, and per‐nucleus intensity features were extracted with CellProfiler. Image‐level separation was assessed by partitioning nuclei into high‐ and low‐intensity groups via Isolation Forest. Statistical comparisons used Welch's t‐tests and Spearman's rank correlation.ResultsThe AC‐30 group included 183 images (AC‐1, n = 183; AC‐2, n = 207). Within AC‐30, 57.4% had non‐AID or unknown diagnoses, 26.8% had other autoimmune diseases, and 15.8% had ANA‐associated rheumatic disease (AARD). In ENA‐negative AC‐30, AARD accounted for 7.7% (6/78), compared to 18.3% (21/115) in ENA‐negative AC‐2. RA accounted for 16.9% (31/183) in AC‐30, consistent across ENA strata. AC‐30 displayed lower per‐nucleus intensities than AC‐2 (all p < 0.0001) and reduced per‐image ΔMaxIntensity (0.177 vs. 0.252, p < 0.0001), while ΔMeanIntensity was similar (0.068 vs. 0.067, p = 0.549).ConclusionsAC‐30 is quantitatively dimmer than AC‐2, with reduced nuclear brightness and peak‐intensity separation. In this retrospective cohort, ENA‐negative AC‐30 was associated with fewer AARD cases. RA accounted for ~17% of AC‐30, and anti‐CCP and RF remained informative markers. The ilastik–CellProfiler workflow enables auditable ANA quantification.Keywords: antinuclear antibody, CellProfiler, HEp‐2, ilastik, image quantification, international consensus on antinuclear antibody patterns, pattern 30, rheumatoid arthritis Introduction Interpretation of antinuclear antibody (ANA) indirect immunofluorescence (IIF) on HEp‐2 cells remains a routine component of autoimmune serology because the titer and, importantly, the staining pattern can provide clinically actionable information beyond a positive or negative screening result [1, 2]. To reduce inter‐laboratory variability and standardize reporting, international harmonization initiatives such as the International Consensus on ANA Patterns (ICAP) have established a common nomenclature and continue to refine pattern definitions as new or under‐characterized patterns are recognized [3, 4]. Within this framework, AC‐30 is defined by a fine speckled nuclear texture in interphase cells accompanied by staining of the metaphase chromatin plate, and it is regarded as an expert‐level nuclear pattern [4, 5]. However, AC‐30 remains primarily defined by qualitative morphology, and standardized quantitative image‐level descriptors that objectively capture its nuclear texture and metaphase plate reactivity are not yet established [5]. The absence of quantitative descriptors can limit reproducibility in routine practice, particularly when fluorescence intensity is modest or when nuclear appearances show overlapping features, thereby motivating objective measurements that can complement expert pattern calls and support consistent reporting [1, 5]. In a large retrospective reassessment, 147 samples corresponding to 8.78% of positive ANA tests were identified as AC‐30 upon re‐evaluation, indicating that AC‐30 is not rare and highlighting the practical value of reproducible descriptors for this pattern [6]. At the same time, the clinical relevance of AC‐30 is still being defined, with published data showing heterogeneous diagnostic contexts and a substantial proportion of AC‐30 cases lacking extractable nuclear antigen (ENA) reactivity, which limits confident disease‐specific interpretation in many settings at present [6]. Although computer‐aided ANA IIF systems can improve throughput and workflow harmonization, evaluations emphasize reduced performance for nuanced or mixed nuclear patterns, and specific evidence for AC‐30 handling remains limited because many platforms were developed and validated before AC‐30 was formally incorporated into ICAP resources [7, 8, 9]. Open‐source bioimage analysis tools provide an auditable alternative for quantitative characterization: Ilastik enables interactive pixel‐level classification using sparse annotations, and CellProfiler enables modular extraction of per‐nucleus features and scalable measurement pipelines [10, 11, 12]. In this study, we aimed to quantitatively characterize AC‐30 using reproducible image‐level features. AC‐30 images were identified by blinded expert rescoring using a predefined consensus rule, and an ilastik–CellProfiler workflow was applied to derive per‐nucleus intensity metrics and image‐level indices for quantitative comparison with AC‐1 and AC‐2, as well as for exploratory associations with clinical and serologic features [10, 11, 12]. Methods Study Design and Oversight This was a retrospective, single‐center analysis of archived HEp‐2 ANA indirect immunofluorescence (IIF) images. Images and linked clinical labels were de‐identified prior to analysis. We screened 82,924 ANA IIF tests performed at the Department of Laboratory Medicine, Peking University Shenzhen Hospital, between February 2022 and June 2025. Inclusion criteria required a stored digital image of adequate technical quality and a readable pattern assignment by expert review. Exclusion criteria were unreadable images, imaging artifacts that prevented pattern interpretation, or missing clinical labels. This study comprised two phases (Figure 1A). Phase I focused on AC‐30 classification and clinical profiling: Archived HEp‐2 ANA indirect immunofluorescence images and linked clinical records were retrieved, AC‐30 images were identified by blinded expert rescoring using a predefined consensus rule, and AC‐30, AC‐1, and AC‐2 cohorts were assembled and summarized by diagnostic categories and available serologic profiles. Phase II focused on quantitative image analysis and clinic‐quantitative associations: An auditable pipeline was developed using ilastik for nucleus segmentation and CellProfiler for per‐nucleus feature extraction, then applied to AC‐30 and comparator images to derive per‐nucleus intensity metrics and predefined image‐level separation indices for prespecified group comparisons; finally, within AC‐30, subgroup analyses and exploratory associations with clinical variables were evaluated, including rheumatoid arthritis vs. other diagnoses and correlations with anti‐CCP where available. Study overview and representative HEp‐2 ANA IIF patterns. A. Two‐phase study schematic. Phase I (blue) covers retrospective image/record retrieval, blinded expert rescoring to define AC‐30, and cohort assembly with AC‐1 and AC‐2 comparators for clinical/serologic profiling. Phase II (orange) covers development of an auditable ilastik‐CellProfiler quantification pipeline, application to derive per‐nucleus and image‐level separation metrics, and subgroup/clinic‐quantitative analyses within AC‐30. B. Representative HEp‐2 ANA IIF images, zoomed and cropped from the original fields of view, showing AC‐1, AC‐2, and AC‐30 patterns selected during expert rescoring and re‐evaluation for cohort definition and comparator selection. Imaging Platform and Acquisition The ANA test was performed by EUROIMMUN HEp‐2 IIFT kit (FC 1510–2010‐1). Slides were processed on the EUROIMMUN Sprinter XL processors at a 1:100 dilution according to the manufacturer's protocol. Image acquisition and evaluation were performed on the EUROPattern Microscope. Exposure and gain settings were kept constant within each batch; raw images were exported as 8‐bit TIFF files without gamma adjustment, and batch identifiers were retained for sensitivity analyses. Blinded Expert Scoring and AC‐30 Identification A total of 82,924 ANA images were independently reviewed by three experienced experts who were blinded to all clinical information. Each expert scored every image on a 1 to 5 scale against predefined visual criteria for the newly codified AC‐30 pattern. Images receiving a unanimous score of 5/5 were classified as AC‐30. For the AC‐1 and AC‐2 comparator cohorts, images were also assessed by three independent experts, and those meeting the respective pattern criteria were randomly selected and frequency‐matched to the AC‐30 cohort based on sample period and image quality. Clinical Labels and Serology Final clinical diagnoses were assessed by the physician according to the national guideline in China and retrieved from the hospital information system (HIS) at the time of ANA reporting and harmonized into predefined categories. For cases with available serology, extractable nuclear antigen (ENA) line immunoassay was performed using the EUROIMMUN anti‐nuclear antibodies profile kit (DL 1590–6401‐3 G) in the routine laboratory. Band intensities were interpreted semi‐quantitatively with EUROLineMaster II as +, ++, or +++. Anti‐cyclic citrullinated peptide (anti‐CCP) was measured using the Anti‐CCP Reagent kit (Abbott) on the Alinity analyzer, and results were retrieved from the HIS. Pixel Classification With Ilastik The ilastik Pixel Classification workflow was used to generate nucleus probability maps from raw ANA IIF images [10]. Images that were compressed, damaged, or differed from the native Sprinter XL output were excluded from quantitative analysis. Images were loaded as 8‐bit TIFF with axis order cyx. Two classes were created: Nucleus and background. Sparse brush labels were drawn on representative fields covering the full intensity range and including AC‐1, AC‐2, and AC‐30. Features were computed in 2D at multiple spatial scales, including Gaussian smoothing, Laplacian ofGaussian, Gaussian gradient magnitude, difference ofGaussians, structure tensor eigenvalues, and Hessian eigenvalues. A random‐forest classifier was trained interactively until the live probability preview stabilized on held‐out fields. The trained project was run in batch to export probability maps as TIFF files. Maps were thresholded at 0.5 to create binary masks, then cleaned by small‐object removal and hole filling. Final masks were saved as TIFF and used as inputs for CellProfiler analysis. Object Segmentation and Measurement With CellProfiler CellProfiler (version 4.2.8) was used to process each image pair using a standardized pipeline [12, 13]. NamesAndTypes module loaded three streams: ORIGIN, nucleus, and cell body. The ColorToGray module first splits the cell body to produce a nucleus‐segmented image and combined ORIGIN into a grayscale image OriGray, with the green channel weight set to 1.0. The IdentifyPrimaryObjects module segmented Nuclei on nucleus‐segmented with a typical diameter of 25–100 pixels, with size filtering enabled, border objects retained, shape‐based declumping, a smoothing filter of 10 pixels, a local maxima distance of 10 pixels, and a Global Otsu two‐class threshold with correction factor 1.0 and a cap of 500 objects per field. The MeasureObjectIntensity module quantified per‐nucleus mean, median, maximum, and edge intensity on OriGray. The OverlayOutlines module produced OrigOverlay for visual quality control, and ExportToSpreadsheet wrote per‐object and per‐image tables as comma‐separated files with image and cohort identifiers. Per‐Image Separation Metrics Within each image, nuclei were partitioned into high and low by fitting an Isolation Forest (scikit‐learn IsolationForest) to log1p(MeanIntensity [OriGray]) from that image only. Settings were set as contamination = 0.05, n_estimators = 25, max_samples = min(128, n), bootstrap = False, random_state = 0; non‐finite values were removed. Nuclei predicted as anomalies (−1) were taken as high only if their raw MeanIntensity was greater than the image median. If no nuclei met this criterion, a fallback strategy selected the top 5% of nuclei by mean intensity; if this still yielded an empty set, the top 1% was used. The remaining nuclei formed the low‐intensity group. Images with fewer than 10 valid nuclei were excluded. For each image, we computed ΔMeanIntensity and ΔMaxIntensity as mean(high) − mean(low) and also reported the corresponding high/low means. All parameters and thresholds were fixed a priori and applied uniformly. Statistical Analysis Analyses were prespecified. Group comparisons of per‐nucleus intensities between AC‐2 and AC‐30 used Welch's t‐tests (two‐sided, α = 0.05) to accommodate unequal variances and sample sizes. For image‐level metrics, Welch's t‐tests compared ΔMeanIntensity and ΔMaxIntensity between AC‐2 and AC‐30. Within AC‐30, RA vs. others was evaluated with the same framework at both the nucleus level and the image level. Correlations between anti‐CCP and image metrics used Spearman rank correlation. No multiplicity adjustment was applied because the three per‐nucleus metrics were defined a priori as co‐primary descriptors of fluorescence intensity. Summary means and counts are reported alongside p‐values where relevant. Statistical computing was performed in Python 3.11 and R 4.3 with standard libraries, including scikit‐learn's IsolationForest for per‐image grouping. Result Blinded Selection of AC‐30 Images and Cohort Characteristics A pooled set of HEp‐2 ANA IIF images (n = 82,924) was independently scored on a 1–5 scale by three experienced readers using predefined criteria; discrepant scores were re‐evaluated according to the adjudication plan. Images with unanimous 5/5 scores were classified as AC‐30 (n = 183) (Figure 1B and Figure 2A). Morphologically, AC‐30 was considered within the dense fine speckled spectrum and described by an extremely fine, grainy nuclear texture with staining of metaphase chromatin, warranting careful distinction from AC‐1 and AC‐2 during classification and interpretation. Comparator cohorts included AC‐1 (n = 183) and AC‐2 (n = 207). Age‐sex distributions in 10‐year bins are shown (Figure 2B) and summarized (Table S1). Independent AC‐30 identification and serologic and clinical stratification across ANA patterns. A. AC‐30 identification workflow. Three experienced ANA readers independently scored pooled HEp‐2 images [1–5]; images with unanimous 5/5 scores were classified as AC‐30 (n = 183). B. Age‐sex population pyramids for AC‐1 (n = 183), AC‐2 (n = 207), and AC‐30 (n = 183) in 10‐year age bins (male left; female right). C. Record‐level ENA line‐immunoassay matrix for AC‐1, AC‐2, and AC‐30 (each row represents one record). White indicates ENA not available; tested results are encoded in grayscale for negative and semi‐quantitative positivity (+, ++, +++). Summary counts indicate records with any ENA positivity (N_pos), all‐negative ENA (N_neg), and no ENA record (N_N.A.). D. Diagnostic distribution for the full AC‐30 cohort (n = 183) and AC‐30 subsets stratified by ENA status (ENA all negative, n = 78; ENA any positive, n = 33). Stacked bars show ANA‐associated rheumatic disease (AARD), other autoimmune disease (other AID), and non‐AID or other or unknown; red annotations indicate the rheumatoid arthritis (RA) proportion. E. Diagnostic distribution for ENA‐defined reference subsets: AC‐1 records with anti‐nucleosome or anti‐histone positivity (n = 31), AC‐2 records with ENA all negative (n = 115), and AC‐30 records with ENA all negative (n = 78). Stacked bars and RA annotations are as in panel D. F. ENA matrix for AC‐30 records with any ENA positivity (n = 33); the rightmost column indicates final diagnosis category (AARD, RA, other AID, or others). ENA testing was performed in routine practice using a fixed line‐immunoassay panel run in parallel with HEp‐2 ANA testing, including anti‐RNP/Sm, anti‐Sm, anti‐SSA, anti‐Ro52, anti‐SSB, anti‐Scl70, anti‐PM‐Scl, anti‐Jo1, ACA, anti‐PCNA, anti‐nucleosome, anti‐histone, anti‐ribosomal P, and anti‐AMA‐M2; no additional ENA specificities were available in this dataset. ENA results across AC‐1, AC‐2, and AC‐30 are summarized (Figure 2C; Table S1). Within the total AC‐30 cohort (n = 183), clinical categories were distributed as non‐AID or unknown 57.4% (105/183), other autoimmune disease 26.8% (49/183), and ANA‐associated rheumatic disease 15.8% (29/183) (Figure 2D). Given that AC‐30‐specific target antigens have not been established, AC‐30 records were further stratified by ENA status to describe clinical profiles with and without concurrent ENA positivity. Among AC‐30 records with ENA all negative (n = 78), diagnoses were non‐AID or unknown 60.0% (47/78), other autoimmune disease 32.1% (25/78), and AARD 7.7% (6/78). Among AC‐30 records with any ENA positivity (n = 33), diagnoses were non‐AID or unknown 45.4% (15/33), other autoimmune disease 24.2% (8/33), and AARD 30.3% (10/33) (Figure 2D). The distribution of ENA specificities and final diagnosis categories within the ENA‐positive AC‐30 subset is shown (Figure 2F). For ENA‐defined reference subsets, AC‐1 records with anti‐nucleosome or anti‐histone positivity (n = 31) were classified as AARD 74.2% (23/31), non‐AID or unknown 22.6% (7/31), and other autoimmune disease 3.2% (1/31) (Figure 2E). AC‐2 records with ENA all negative (n = 115) were classified as non‐AID or unknown 68.7% (79/115), other autoimmune disease 13.0% (15/115), and AARD 18.3% (21/115) (Figure 2E). Rheumatoid Arthritis Enrichment Within AC‐30 and Associated Serologic Profile Within the AC‐30 cohort (n = 183) and its ENA‐stratified subsets, the proportion of rheumatoid arthritis (RA) was similar across groups, accounting for 16.9% (31/183) in the total AC‐30 cohort, 16.7% (13/78) in AC‐30 records with ENA all negative, and 15.2% (5/33) in AC‐30 records with any ENA positivity (Figure 2D). In ENA‐defined reference subsets, RA was less frequent in AC‐1 records with anti‐nucleosome or anti‐histone positivity (3.2%, 1/31) and in AC‐2 records with ENA all negative (7.8%, 9/115) than in AC‐30 records with ENA all negative (16.7%, 13/78) (Figure 2E). Clinical characteristics and autoantibody testing results stratified by diagnostic category within AC‐30 are summarized (Table 1). Female proportion and age did not differ between RA and AARD (female 87.1% vs. 100.0%, p = 0.114; age median 45.0 vs. 49.5 years, p = 0.514), and were also comparable between RA and the Others group (female 87.1% vs. 73.4%, p = 0.157; age median 45.0 vs. 53.0 years, p = 0.051). As expected, anti‐CCP levels were markedly higher in RA than in AARD and Others (RA 154.5 RU/mL [45.4–599.6], n = 18; AARD 0.5 [0.5–0.6], n = 7; Others 0.5 [0.5–0.5], n = 15; both p < 0.001). RF levels were also higher in RA than in AARD and Others (RA 117.0 IU/mL [42.8–337.5], n = 25; AARD 10.0 [10.0–10.0], n = 12; Others 10.0 [10.0–10.0], n = 27; both p < 0.001). In contrast, ESR and hs‐CRP were numerically higher in RA than in AARD and Others, but the differences were not statistically significant (ESR: RA 20.0 mm/h [9.5–34.5], n = 27; AARD 18.0 [6.0–28.0], n = 17; Others 14.0 [8.0–32.0], n = 43; p = 0.531 and p = 0.282; hs‐CRP: RA 1.40 mg/L [0.93–10.04], n = 21; AARD 0.72 [0.56–1.72], n = 8; Others 1.28 [0.67–2.51], n = 17; p = 0.093 and p = 0.290) (Table 1). TABLE 1Clinical characteristics and autoantibody testing results stratified by diagnostic category within the AC‐30 cohort.VariableRAAARD P (AARD vs. RA)Others P (others vs. RA)N3128124Female sex, n (%) a 27 (87.1%)28 (100.0%)0.11491 (73.4%)0.157Age (years) b : median (IQR) [n]45.0 (32.0–56.5) [n = 31]49.5 (36.8–58.5) [n = 28]0.51453.0 (39.0–65.5) [n = 123]0.051ENA: Not performed, n (%)12 (38.7%)11 (39.3%)49 (39.5%)ENA: All negative, n (%)14 (45.2%)7 (25.0%)57 (46.0%)ENA: Positive, n (%)5 (16.1%)10 (35.7%)18 (14.5%)CCP (RU/mL) b : median (IQR) [n]154.5 (45.4–599.6) [n = 18]0.5 (0.5–0.6) [n = 7]< 0.0010.5 (0.5–0.5) [n = 15]< 0.001ESR (mm/h) b : median (IQR) [n]20.0 (9.5–34.5) [n = 27]18.0 (6.0–28.0) [n = 17]0.53114.0 (8.0–32.0) [n = 43]0.282RF (IU/mL) b : median (IQR) [n]117.0 (42.8–337.5) [n = 25]10.0 (10.0–10.0) [n = 12]< 0.00110.0 (10.0–10.0) [n = 27]< 0.001hs‐CRP (mg/L) b : median (IQR) [n]1.40 (0.93–10.04) [n = 21]0.72 (0.56–1.72) [n = 8]0.0931.28 (0.67–2.51) [n = 17]0.290 Note: p values were calculated for pairwise comparisons vs. the RA group. a Female sex was compared using a two‐sided Fisher's exact test. b Age and laboratory variables, including CCP (U/mL), ESR (mm/h), RF (IU/mL), and hs‐CRP (mg/L), were compared using a two‐sided Mann–Whitney U test and are presented as median (IQR) with the number of non‐missing observations [n]. Missing values were excluded from the corresponding analyses. For laboratory results reported with “<” or “>”, the reported threshold value was used for statistical testing, while the original text value was retained in the dataset. Clinical characteristics and autoantibody testing results stratified by diagnostic category within the AC‐30 cohort. Note: p values were calculated for pairwise comparisons vs. the RA group. Female sex was compared using a two‐sided Fisher's exact test. Age and laboratory variables, including CCP (U/mL), ESR (mm/h), RF (IU/mL), and hs‐CRP (mg/L), were compared using a two‐sided Mann–Whitney U test and are presented as median (IQR) with the number of non‐missing observations [n]. Missing values were excluded from the corresponding analyses. For laboratory results reported with “<” or “>”, the reported threshold value was used for statistical testing, while the original text value was retained in the dataset. Develop an Auditable Workflow for Quantitative Analysis of ANA Results Using Ilastik and CellProfiler An auditable two‐stage workflow was developed and implemented to quantify ANA results at single‐nucleus resolution (Figure 3). In the first stage, ilastik was used for pixel classification to separate nuclei from background, generating probability maps and binary masks for batch processing (Figure 3A). In the second stage, we used CellProfiler to import images and masks, segment individual nuclei with the IdentifyPrimaryObjects module, overlay outlines for quality control, and extract intensity features (MeanIntensity, MedianIntensity, MaxIntensity, EdgeIntensity) with the MeasureObjectIntensity module. Resulting feature tables were exported for statistical analysis (Figure 3B, Table S2). Given that AC‐30 has been primarily described morphologically, this workflow is presented as a reproducible framework to complement descriptive assessment without asserting a definitive standard. Ilastik and CellProfiler workflow for ANA image quantification. A. Pixel classification and batch segmentation in ilastik. A classifier was trained on representative HEp‐2 ANA images to distinguish nuclei from background, producing probability maps and binary masks that were applied to the full image set in batch. B. CellProfiler measurement pipeline. Images and masks were imported, converted to grayscale, and nuclei were segmented with IdentifyPrimaryObjects. Outlines were overlaid for quality control, and per‐nucleus intensity features were extracted with MeasureObjectIntensity (e.g., mean, median, maximum, and edge intensity) and exported as tables. AC‐30 Exhibits Lower Per‐Nucleus Intensities Than AC‐2 and Reduced Per‐Image Peak Separation Application of the workflow to AC‐2 and AC‐30 confirmed accurate nucleus detection across patterns (Figure 4A). After quality control, 76 AC‐2 images (yielding 30,778 nuclei) and 97 AC‐30 images (yielding 48,787 nuclei) were successfully analyzed (Table S2). At the single‐nucleus level, AC‐30 was dimmer than AC‐2 across all metrics (Figure 4B–D): MeanIntensity 0.253 vs. 0.281, MedianIntensity 0.258 vs. 0.296, MaxIntensity 0.389 vs. 0.483 (all p < 0.0001). To summarize within‐image structure, nuclei in each image were partitioned into high‐ and low‐intensity groups using Isolation Forest on MeanIntensity, and per‐image separation was defined as Δ = mean(high) − mean(low) (Figure 4E). ΔMeanIntensity was similar in AC‐2 and AC‐30 (0.067 vs. 0.068, p = 0.549), whereas ΔMaxIntensity was smaller in AC‐30 (0.177 vs. 0.252, p < 0.0001) (Figure 4F,G, and Table S2). These findings support a globally lower signal in AC‐30 and a reduced separation of peak intensities at the image level. Quantification of AC‐2 and AC‐30 intensity features from ANA images. A. Pipeline applied to AC‐2 and AC‐30. Ilastik pixel classification produces masks; CellProfiler IdentifyPrimaryObjects detects nuclei for per‐nucleus intensity measurement. B−D. Beeswarm plots of all nuclei. B. MeanIntensity, C. MedianIntensity, D. MaxIntensity. AC‐2 n = 30,788 and AC‐30 n = 48,787; group means: 0.281 vs. 0.253, 0.296 vs. 0.258, and 0.483 vs. 0.389, respectively. All comparisons by Welch t‐test; p < 0.0001. E. Per‐image workflow. For each image, Isolation Forest on MeanIntensity separates high‐ and low‐intensity nuclei; Δ is computed as mean(high) − mean(low) for MeanIntensity and MaxIntensity. F. Beeswarm of per‐image ΔMeanIntensity for AC‐2 (N = 76) and AC‐30 (N = 97); Welch p = 0.549; means 0.067 vs. 0.068. G. Beeswarm of per‐image ΔMaxIntensity for AC‐2 (N = 76) and AC‐30 (N = 97); Welch p < 0.0001; means 0.252 vs. 0.177. Quantitative Analysis Shows No Image Level Differences by ENA Status or Diagnosis, With Lower Nucleus Level Intensity in Rheumatoid Arthritis Among the 97 quantitatively analyzed AC‐30 images (Table S2), 94 had definitive diagnoses in the hospital information system, including 19 rheumatoid arthritis cases (9,207 nuclei) and 75 other diagnoses (37,743 nuclei) (Table S2 and Table S3). ENA results were available for a subset of these images, including 35 ENA all‐negative and 13 ENA‐positive records (Table S3). To assess whether concurrent ENA reactivity was associated with differences in AC‐30 image, AC‐30 images were stratified by ENA status and evaluated using per‐image analyses after Isolation Forest splitting of nuclei into high‐ and low‐intensity groups. Image‐level separation did not differ between ENA all‐negative (N = 35) and ENA‐positive (N = 13) AC‐30 images for either ΔMeanIntensity (p = 0.884) or ΔMaxIntensity (p = 0.93) (Figure 5A,B). AC‐30 image quantitative features by ENA status and rheumatoid arthritis subgroup. A, B. Per‐image analyses after Isolation Forest splitting of nuclei into high‐ and low‐intensity groups in AC‐30 images stratified by ENA status (ENA all negative N = 35; ENA any positive N = 13). Separation was defined as Δ(high‐low) = mean(high) − mean(low). A. ΔMeanIntensity. B. ΔMaxIntensity. Group means are indicated; p values are from Welch's t‐test. C. Spearman correlations between anti‐CCP level and per‐image intensity metrics in AC‐30 rheumatoid arthritis cases with available serology (N = 8); bars show correlation coefficients, and p values are annotated. D, E. Per‐image analyses after Isolation Forest splitting comparing AC‐30 images from rheumatoid arthritis (N = 19) vs. other diagnoses (N = 75). D. ΔMeanIntensity. E. ΔMaxIntensity. Group means are indicated; p values are from Welch's t‐test. F−H. Per‐nucleus intensity distributions within AC‐30 comparing rheumatoid arthritis vs. other diagnoses. Beeswarm plots show all nuclei measured by CellProfiler, with group means annotated. F. MeanIntensity, G. MedianIntensity, and H. MaxIntensity. Group comparisons use Welch's t‐test. AC‐30 images were also analyzed by diagnostic subgroup. In the RA subset with available serology (N = 8), anti‐CCP levels showed no correlation with per‐image quantitative metrics (all p > 0.49, Spearman) (Figure 5C; Table S3). Consistently, image‐level separation metrics did not differ between RA images (N = 19) and images from other diagnoses (N = 75) for ΔMeanIntensity (p = 0.263) or ΔMaxIntensity (p = 0.645) (Figure 5D,E, Table S2). In contrast, nucleus‐level features differed by diagnosis. Within AC‐30, nuclei from RA patients showed lower per‐nucleus intensity than nuclei from other diagnoses across MeanIntensity, MedianIntensity, and MaxIntensity (all p < 0.0001, Welch's t‐test) (Figure 5F–H; Table S2). The largest absolute difference was observed for MaxIntensity (RA mean 0.259 vs. other diagnoses 0.273) (Figure 5H). Discussion AC‐30 is a recently codified ICAP ANA pattern for which quantitative image‐level descriptors and clinical relevance have not been fully established; in this study, we address this gap through systematic image quantification [5]. In this retrospective cohort, ENA stratification suggested different clinical profiles within AC‐30: The ENA all‐negative AC‐30 subset showed a low proportion of AARD, 7.7% (6/78), whereas AARD was more frequent among ENA‐positive AC‐30 records, 30.3% (10/33) (Figure 2D). In ENA‐defined reference subsets, the ENA all‐negative AC‐30 subset showed a lower AARD proportion, 7.7% (6/78), than the ENA all‐negative AC‐2 reference subset, 18.3% (21/115) (Figure 2E). These distributions may be influenced by retrospective sampling and incomplete serology, but they raise the possibility that ENA all‐negative AC‐30 in routine testing may be enriched for non‐AARD presentations; however, AC‐30 pattern and intensity features are now not sufficient to infer a specific diagnosis in individual cases and should be interpreted with clinical context and disease‐specific serology. These findings are of interest in the context of prior work on AC‐2 and anti‐DFS70. The AC‐2 dense fine speckled pattern is often linked to anti‐DFS70 antibodies, and monospecific anti‐DFS70 has been reported to help lower the likelihood of AARD when other disease‐associated antibodies are absent, including meta‐analytic evidence and earlier clinical summaries [14, 15, 16]. Prospective studies with standardized clinical adjudication and expanded serology are further needed to determine whether ENA all‐negative AC‐30 can be interpreted similarly to “isolated” DFS70‐associated patterns in reducing the likelihood of AARD, or whether this reflects center‐specific testing patterns. Notably, DFS70 was not tested in this retrospective dataset; a recent report suggested that relatively lower anti‐DFS70 levels or coexistence with a high‐titer homogeneous pattern may favor an AC‐30 appearance rather than AC‐2 [17]. Rheumatoid arthritis appeared relatively enriched within AC‐30 compared with ENA‐defined reference subsets and was not explained by ENA status. RA accounted for 16.9% (31/183) of the total AC‐30 cohort and remained similar across ENA‐stratified AC‐30 subsets (Figure 2D), whereas RA was less frequent in ENA‐defined reference subsets of AC‐1 and AC‐2 (Figure 2E). Within AC‐30, RA remained clinically supported by routine serology, with significantly higher anti‐CCP and RF levels than in AARD and in Other diagnoses (Table 1). These results suggest that AC‐30 encountered in RA should still be interpreted alongside established RA serology rather than implying a distinct RA‐specific ANA signature. A major contribution of this work is the development and application of an auditable quantitative framework for ANA image analysis using widely available open‐source tools. Ilastik and CellProfiler enabled transparent nucleus segmentation and feature extraction with fully specified parameters, supporting reproducibility across laboratories with similar imaging conditions. Quantitatively, AC‐30 was dimmer than AC‐2 at the nucleus level (mean intensity 0.253 vs. 0.281, p < 0.0001) and showed reduced image‐level peak separation (ΔMaxIntensity 0.177 vs. 0.252, p < 0.0001), while ΔMeanIntensity was similar (0.068 vs. 0.067, p = 0.549) (Figure 4). These measurements provide objective descriptors for distinguishing AC‐30 from AC‐2 in borderline cases, consistent with expert descriptions of AC‐30 as an extremely fine, low‐contrast pattern within the dense fine speckled spectrum [5]. Within AC‐30, quantitative subgroup signals were detectable at the nucleus level but did not extend to image‐level structure. ENA status was not associated with differences in image‐level separation (Figure 5A,B), and image‐level separation also did not differ between RA and other diagnoses (Figure 5D,E); anti‐CCP showed no correlation with per‐image quantitative metrics in the small RA subset with available serology (Figure 5C). In contrast, nuclei from RA cases showed lower per‐nucleus intensities across MeanIntensity, MedianIntensity, and MaxIntensity (all p < 0.0001), with the largest absolute difference observed for MaxIntensity (0.259 vs. 0.273) (Figure 5F–H). Although this nucleus‐level difference was statistically significant, the effect size was modest and should be interpreted cautiously given the retrospective design, unequal subgroup sizes, and the absence of corresponding image‐level separation. Notably, RA cases in the AC‐30 cohort remained clinically supported by conventional RA serology (Table 1), consistent with contemporary RA classification frameworks that incorporate these markers [18, 19]. Further studies with expanded antibody testing and prospective sampling are needed to determine whether these subtle nucleus‐level differences are reproducible and clinically meaningful [20]. Several limitations should be noted. This was a single‐center retrospective study based on archived images and routine clinical records, which may introduce selection and information bias. Serology was limited to a fixed ENA line‐immunoassay panel and did not include DFS70 or anti‐dsDNA in the available dataset, restricting mechanistic inference and limiting comparability to established DFS70‐focused frameworks. Broader standardization of capture settings, multi‐reader consensus datasets, and prospective validation will be important next steps to determine whether intensity‐based descriptors improve agreement and reduce misclassification for this recently codified pattern. Author Contributions Changmeng Wu and Yu Li conceived and designed the study; Changmeng Wu and Wenhan Du developed the ilastik‐CellProfiler pipeline and performed formal analyses; Linlin Xu and Juan He curated clinical and serological data; Changmeng Wu prepared figures and tables and drafted the manuscript; Wenhan Du, Weidong Chen, Linlin Xu, and Yu Li critically revised the manuscript; Yu Li provided resources, supervision, and project administration. All authors approved the final manuscript and agree to be accountable for all aspects of the work. Funding This work was supported by the National Natural Science Foundation of China (81903069), the Basic and Applied Basic Research Foundation of Guangdong Province (2023A1515010294), the Shenzhen Science and Technology Program (JCYJ20210324110011031), the Shenzhen High‐level Hospital Construction Fund, Medical Scientific Research Foundation of Guangdong Province (B2026745) and the Peking University Shenzhen Hospital Scientific Research Fund. Ethics Statement This retrospective study analyzed de‐identified, archived HEp‐2 ANA images and associated labels only; no patient recruitment or re‐contact occurred. In accordance with institutional and national regulations, the requirement for ethics approval was waived and informed consent was approved by the Ethics Committee of Peking University Shenzhen Hospital. Conflicts of Interest The authors declare no conflicts of interest. Supporting information Table S1: Case listing, diagnostic demographics, and ENA serology by ANA pattern. Table S2: Per‐nucleus measurements and per‐image separation metrics of the AC‐2 and AC‐30 patterns. Table S3: Case listing and clinical information of the 94 quantitatively AC‐30 samples. Acknowledgments This research is supported by the National Natural Science Foundation of China (81,903,069); Basic and Applied Basic Research Foundation of Guangdong Province (2023A1515010294); Shenzhen Science and Technology Program (JCYJ20210324110011031); Shenzhen High‐level Hospital Construction Fund; Peking University Shenzhen Hospital Scientific Research Fund. The authors thank EUROIMMUN Medical Diagnostics (China) Co. Ltd. for assistance with retrospective image‐quality review and technical support. Data Availability Statement The data that supports the findings of this study are available in the Table S1–Table S3 of this article. References Associated Data Supplementary Materials Table S1: Case listing, diagnostic demographics, and ENA serology by ANA pattern. Table S2: Per‐nucleus measurements and per‐image separation metrics of the AC‐2 and AC‐30 patterns. Table S3: Case listing and clinical information of the 94 quantitatively AC‐30 samples. Data Availability Statement The data that supports the findings of this study are available in the Table S1–Table S3 of this article.