Evolving landscape of imaging-based evaluation in systemic autoimmune rheumatic disease-associated interstitial lung disease: from visual assessment to quantitative artificial intelligence-assisted evaluation

Abstract

Interstitial lung disease (ILD) is a major driver of morbidity and mortality across systemic autoimmune rheumatic diseases (SARDs), with systemic sclerosis–associated ILD (SSc-ILD) providing the most extensive evidence base. In this context, progressive pulmonary fibrosis has emerged as a central framework, as it is associated with increased mortality and facilitates the identification of candidates for antifibrotic therapy. Nevertheless, operational thresholds for chest high-resolution computed tomography (HRCT)–defined progression remain ill defined: current guidelines rely on visual HRCT interpretation and lack standardized, reproducible assessment protocols. Because the magnitude and topography of disease evolution can guide therapeutic decisions, quantitative evaluation of imaging features is pivotal. In this review, we delineate the evolution of imaging assessment from qualitative reads to quantitative phenotyping. We organize traditional densitometric and textural metrics (e.g., percentage high-attenuation areas, quantitative lung fibrosis, CALIPER [Computer-Aided Lung Informatics for Pathology Evaluation and Ratings]) alongside hybrid/data-driven approaches (e.g., data-driven textural analysis, quantitative interstitial abnormality) and recent deep-learning tools (e.g., SOFIA [Systemic Objective Fibrotic Imaging Analysis Algorithm], eLung, Qureight, SATORI [Segmentation and Annotation Tool for Radiomics and Deep Learning], AirQuant). Given the rapid pace of innovation in artificial-intelligence–based quantitative CT, we present a curated set of analytic approaches and offer a concise framework for understanding technological progress and evaluating its relevance to SARD-ILD applications.

INTRODUCTION

Interstitial lung disease (ILD) is a major contributor to morbidity and mortality in patients with systemic autoimmune rheumatic diseases (SARDs). A systematic review estimated the pooled prevalence of ILD at 40%~50% in systemic sclerosis (SSc) and idiopathic inflammatory myositis (IIM) [1], with mortality approximately twice as high in both conditions [2,3]. Similarly, ILD occurs in up to 20% of patients with rheumatoid arthritis (RA) [4,5], and is associated with more than two-fold increased mortality [6-12]. Because the clinical trajectory of SARD-ILD is heterogeneous [13-16], stratifying risk and recognizing patients likely to progress are central to improving outcomes. Current evidence highlights ILD progression as a central factor associated with poor outcomes in this population [2,8,17,18].

Progression, however, lacks a uniform definition in SARD-ILD, leading to the application of varied criteria across studies [2,8,15,17-19]. Many of these frameworks derive from research in fibrosing ILDs such as idiopathic pulmonary fibrosis (IPF). Most definitions rely either exclusively on pulmonary function tests (PFTs) or on a composite of PFTs, high-resolution computed tomography (HRCT), and respiratory symptoms [15,20,21]. Decline in forced vital capacity (FVC), with or without concurrent reduction in diffusing capacity for carbon monoxide (DLCO), is used most commonly and serves as the primary surrogate endpoint in ILD trials because of its robust association with mortality [22]. However, FVC is effort-dependent and susceptible to measurement variability, complicating assessment of treatment effects in IPF and related ILDs.

HRCT offers a complementary, protocol-driven evaluation that is less prone to patient effort and may yield more objective, reproducible measures of disease status. Using standardized thin-section (≤1.5 mm) volumetric acquisition with high-spatial-frequency reconstruction, visual inspection of HRCT by structured criteria can identify presence of ILD as well as characteristic imaging patterns such as usual interstitial pneumonia (UIP), probable UIP, nonspecific interstitial pneumonia (NSIP), and organizing pneumonia (OP) that may have prognostic significance. In addition, visual inspection of HRCT can characterize distribution (e.g., basal or subpleural predominance), extent, and pertinent ancillary features (e.g., traction bronchiectasis, emphysema, air trapping, vascular remodeling, nodules, pleural abnormalities, and infection). In SARD-ILD, such image-based assessments contextualize physiological change, refine risk stratification, and inform treatment planning.

Therefore, the next critical step would be to convert granular HRCT observations into standardized, objective outcome measures that are suitable for reproducible assessment. Extensive work has pursued quantitative CT methodologies. This review summarizes recent progress in quantitative HRCT interpretation and delineates future research directions in this rapidly advancing area.

MAIN SUBJECTS

Visual assessment

The first semi-quantitative visual assessment scoring system of SARD-ILD was reported by Warrick et al. [23] based on CT images of SSc-ILD patients. In this scoring system, two main scores—severity and extent score were summed. Severity scores were calculated summing scores of ILD characteristics such as ground-glass opacities (GGOs), irregular pleural margin, septal or subpleural lines, honeycombing (HC), and subpleural cysts, while extent score was measured by counting number of affected lung segments.

Another method for grading CT images was derived from a comparative study that analyzed the relationship between HRCT findings and lung histopathology in patients with SSc-ILD [24]. In this study, HRCTs were graded according to the extent of the dominant pattern (parenchymal opacification vs. reticulation). They demonstrated that predominantly reticular abnormalities on HRCT are strongly associated with a fibrotic histological appearance on lung biopsy. In subsequent research, the authors established a refined semi-quantitative HRCT scoring system and analyzed its association with pulmonary function in SSc-ILD [25,26]. Within this framework, the above-mentioned dominant imaging pattern was quantified semi-quantitatively in 5% increments across five standardized lung levels. The total burden of disease was subsequently derived by combining all individual scores to estimate overall parenchymal involvement. In this study, the precent predicted DLCO best reflects the extent of fibrosing alveolitis in fibrosing alveolitis associated with SSc. This grading approach subsequently evolved into the Goh and Wells criteria, which remain widely referenced in both SSc-ILD and broader SARD-ILD research [27]. The system integrates radiological evaluation from HRCT and pulmonary function assessment. Patients are stratified according to HRCT extent as <20%, indeterminate, or >20%. For indeterminate cases, FVC is used as an additional determinant, with thresholds of ≥70% and <70%. Disease is classified as limited when HRCT extent is <20% or indeterminate with FVC ≥70%, whereas extensive disease is defined as HRCT >20% or indeterminate with FVC <70%. This binary staging system proved to be a robust predictor of mortality, remaining significant after adjustment for confounders such as age, sex, smoking, and pulmonary hypertension (hazard ratio [HR], 3.66; 95% confidence interval [CI], 2.25~5.97; p<0.0005).

To assess the clinical relevance of radiological abnormalities in SSc-ILD, Ooi et al. [28] introduced a lobar-based HRCT visual scoring system that distinguishes between inflammatory and fibrotic changes. HRCT findings—including pure GGO, GGO with reticular-nodular opacities, reticular opacities alone, and HC—were graded for each lobe using a four-level semiquantitative scale (<25%, 25%~49%, 50%~74%, and ≥75%). Scores were aggregated into an inflammatory index (GGO+mixed patterns) and a fibrosis index (reticular+honeycomb patterns). Their analysis revealed that the inflammatory index correlated negatively with DLCO (r=−0.43, p=0.008) and total lung capacity (TLC) (r=−0.39, p<0.05), whereas the fibrosis index correlated with FVC (r=−0.31, p=0.05) and TLC (r=−0.38, p=0.02).

The most widely adopted semiquantitative visual scoring system to date was introduced in the landmark Scleroderma Lung Study I (SLS I) that investigated cyclophosphamide versus placebo [29]. In this system, each lung was divided into three anatomical zones: the upper zone (apex to aortic arch), middle zone (aortic arch to inferior pulmonary veins), and lower zone (inferior pulmonary veins to the lung bases). Within each zone, HRCT abnormalities—including pure GGO, PF, HC, and emphysema—were scored on a 5-point scale: 0, absent; 1, 1%~25%; 2, 26%~50%; 3, 51%~75%; and 4, 76%~100%. This visual scoring approach was employed as the standard outcome measure in both the SLS I and SLS II trials and also served as a comparator against computer-aided quantitative CT scoring methods [30-38].

While most semiquantitative CT scoring systems have been developed in the context of SSc-ILD, Nurmi et al. [39] proposed a system based on RA-ILD. This method separately evaluated the extent of multiple radiological findings, including GGO, reticulation, HC, emphysema, consolidation, crazy-paving appearance, bronchiectasis, traction bronchiectasis, nodules, bronchovascular bundle thickening, cysts, mosaic attenuation, air trapping, rounded atelectasis, architectural distortion, pleural plaques, pleural effusion, and tumors. The lungs were divided into upper, middle, and lower zones, and the extent of GGO, reticulation, and HC was graded on a 0~4 semiquantitative scale. The total score for these three key features was derived by summing all six zones, with a maximum possible value of 24. Their analysis revealed that reticulation (HR, 1.144; p=0.041), traction bronchiectasis (HR, 1.184; p=0.030), architectural distortion (HR, 1.094; p=0.044), and pleural effusion (HR, 14.969; p<0.001) were significant predictors of worse survival. Visual HRCT scoring systems used by expert readers are summarized in Table 1.

Table 1.

Summary of selected visual scoring systems of HRCT by expert readers

Method	Unit of analysis	Grading	Item	Ref
Warrick et al. 1991	Bronchopulmonary segments	Severity score 1. GGOs 2. Irregular pleural margin 3. Septal or subpleural lines 4. HC 5. Subpleural cysts Extent score Bronchopulmonary segments 1. 1~3 segments 2. 4~9 3. >9 segments	GGOs, irregular pleural margin, septal or subpleural lines, HC, subpleural cyst	[23]
Wells et al. 1992	Biopsied lobes	By dominant pattern in extent 1. Parenchymal opacification alone 2. Parenchymal opacifications> reticular pattern 3. Parenchymal opacifications≈ reticular pattern 4. Reticular pattern parenchymal>opacification 5. Reticular pattern alone	Parenchymal opacification and reticular pattern	[24]
Wells et al. 1997 and Collins et al. 1994	Five levels 1. Origin of great vessels 2. Mid-aortic arch 3. Carina 4. Pulmonary venous confluence 5. 1 cm above right diaphragm	Visual estimate of parenchymal abnormality at each level (nearest 5%) total score calculated by mean of 5 levels	Extent of parenchymal abnormality Semi-quantitative evaluation of the predominant radiographic pattern Grade 1: GGO>reticulation Grade 2: GGO=reticulation Grade 3: Reticulation>GGO	[25,26]
Ooi et al. 2003	Lobes are scored independently	0. 0% 1. 1%~25% 2. 26%~50% 3. 51%~75% 4. >75%	Ground-glass opacity alone Mixed groud glass and reticular disease Reticular fibrosis alone HC	[28]
Goldin et al. 2008	3 zones for each lung (upper, mid, lower)	ILD extent: 0~4 scale 0. None 1. 1%~25% 2. 26%~50% 3. 51%~75% 4. >75%	Pure GGO, reticular fibrosis (including thickened reticular markings, bronchiectasis and bronchiolectasis), HC	[29]
Nurmi et al. 2018	3 zones for each lung The upper zones: at or superior to the aortic arch; the middle zones: between the aortic arch and pulmonary veins; the lower zones: at or below the pulmonary veins	0. Absent 1. Minor peripheral/single scattered changes 2. Uniform peripheral or minor central scattered/larger single changes or several minor changes 3. Substantial peripheral changes that penetrated deeply into the lung parenchyma/uniform or substantial changes 4. Very abundant peripheral and central changes	(22 items) Ground-glass opacity HC Reticulation Emphysema Consolitation Crazy-paving appearance Bronchiectasis Traction bronchiectasis Nodules Thickening of bronchovascular bundle Cysts Mosaic attenuation Air trapping Subpleural lines Irregular opacities Subpleural GGO Parenchymal bands Rounded atelectasis Architectural distortion Pleural plaques Pleural effusion Tumor	[39]

HRCT: high-resolution computed tomography, GGOs: ground-glass opacities, HC: honeycombing, ILD: interstitial lung disease.

1) Advantages and limitations of visual semi-quantitative scoring system

Visual scoring systems provide the ability to extract HRCT characteristics systematically to enable understanding of image features and their histological, physical, and clinical outcomes in SARD-ILDs. The major advantages of visual assessment are that it is relatively easily accessible to the clinical field, applicable to routine clinical practice, and does not require any specialized equipment or software. Nevertheless, there has been controversy regarding inter- and intra-observer variability of semi-quantitative scoring, which requires highly experienced and trained chest radiologist to minimize that variability [40,41]. Some research studies require multiple readers and adjudication for disagreements, which can be time-consuming. The “sequential reader” method has been used to facilitate efficiency in visual inspection. In this method, a single reader views all scans, and a second reader reviews all positive scans and only a portion of negative scans [42]. In clinical practice, the structured methods may not be routinely followed, particularly outside specialized centers, where ILD imaging expertise is limited. In addition, visual assessment may systematically underestimate the presence and amount of disease as radiologists often default to a negative score in ambiguous situations. This may be particularly problematic when evaluating interstitial lung abnormalities (ILAs), which often have subtle abnormalities and engender false negatives and higher levels of disagreements due to marginal findings [43]. As a consequence, semi-quantitative methods may miss a chance to capture more sophisticated precise prediction for prognosis.

Quantitative CT assessment

Quantitative CT assessment encompasses diverse approaches, including whole-lung volumetric analysis, density histogram–based evaluation, airway and vascular morphometry, and texture- or radiomics-based pattern recognition. Compared with traditional semiquantitative visual assessment, quantitative CT provides greater objectivity, improved reproducibility, enhanced sensitivity to subtle longitudinal changes, and the potential to serve as a quantitative imaging biomarker in clinical trials. Historically, quantitative CT methods were manually engineered or rule-based, relying on radiology specialists to predefine pathological features and extract relevant measurements (e.g., densitometry, quantitative lung fibrosis [QLF], CALIPER [Computer-Aided Lung Informatics for Pathology Evaluation and Ratings]). More recently, the field has shifted toward artificial intelligence (AI)-driven or data-driven approaches, in which machine learning and deep learning algorithms classify lesions and infer disease states directly from imaging data, often integrated with clinical information. We next review the progression from conventional, feature-engineered quantitative CT methods to next-generation, machine learning–based quantitative CT methods, reflecting the dynamic evolution of quantitative imaging in ILD. Selected quantitative CT methods, categorized by imaging domain and measurement approach, are presented in Table 2. A concise summary of selected quantitative computed tomography (QCT) methods is provided in Table 3.

Table 2.

Selected QCT methods categorized by imaging domain and measurement approach

Method	Parenchymal	Airway	Vascular
Traditional/feature-engineered
Densitometry (%HAA)	-	–	–
Vitrea software	✔ (Density, emphysema, ILD extent)	–	–
QLF	-	–	–
FRI	✔ (Ventilation–perfusion effect)	✔ (Airflow modeling)	✔ (Hemodynamic modeling)
Radiomics	✔ (Texture, density, heterogeneity)	✔ (Airway radiomics)	✔ (Vascular radiomics)
Hybrid
CALIPER	✔ (GGO, reticulation, HC, emphysema)	–	✔ (VRS)
DTA	-	–	–
QIA	✔ (HAA refinement, voxel classifier)	–	–
AirQuant	–	✔ (Airway segmentation, tapering, tortuosity)	–
AI-based/data-driven
eLung	✔ (Fibrosis quantification)	–	–
icolung	✔ (COVID/ILD lesion volume)	✔ (Airway segmentation)	✔ (Vascular involvement)
SOFIA	✔ (DL-based UIP classification)	–	–
SATORI	✔ (CNN-based segmentation of GGO, reticulation)	–	–
Qureight	✔ (Fib module: fibrosis quantification)	✔ (Air module: airway morphometry)	✔ (Vas module: vascular segmentation/volume)

QCT: quantitative computed tomography, HAA: high-attenuation areas, ✔: yes, –: no, ILD: interstitial lung disease, QLF: quantitative lung fibrosis, FRI: functional respiratory imaging, CALIPER: Computer-Aided Lung Informatics for Pathology Evaluation and Ratings, GGOs: ground-glass opacities, HC: honeycombing, VRS: vessel-related structures, DTA: data-driven textural analysis, QIA: quantitative interstitial abnormality, AI: artificial intelligence, COVID: Coronavirus disease, SOFIA: Systemic Objective Fibrotic Imaging Analysis Algorithm, DL: deep learning, UIP: usual interstitial pneumonia, SATORI: Segmentation and Annotation Tool for Radiomics and Deep Learning, CNN: convolutional neural network.

Table 3.

Summary of selected QCT methods

Method	Primary purpose	Summary methods	Outcome metrics	Ref
Traditional/feature-engineered
Densitometry	Detect high-density voxels suggestive of subclinical ILD/inflammation	Threshold-based histogram; voxels in HU range (e.g., −600 to −250 HU)	% of lung volume in HAA range	[45-47]
Vitre software	Commercial densitometry (clinical use)	Automated lung density & volume quantification	%LAA, %HAA, lung volume	[48-50]
QLF	Quantify fibrosis extent for ILD severity	Semi-automated texture analysis; rule-based scoring of GGO, reticulation, HC	%fibrosis involvement; severity index	[36-36,38,55-55]
FRI	Integrative airway and functional modeling	3D reconstruction of airway/vasculature with airflow simulation	Regional airway resistance, functional indices	[58-58,60]
Radiomics (classical)	Feature-based ILD characterization	High-dimensional texture, density, heterogeneity features	Radiomic risk score; heterogeneity metrics	[61-69]
Hybrid
CALIPER	ILD pattern classification & quantification; also vascular metrics (VRS)	ML-based voxel classification (normal, GGO, reticulation, HC, emphysema); VRS	%volume per ILD subtype; %VRS	[70-79]
DTA	Quantify ILD patterns using unsupervised clustering	Histogram- and texture-based feature grouping	%abnormal lung volume	[80-82]
QIA	Refined quantification of interstitial changes	HAA refinement+ML voxel classifier	%QIA extent	[83-86]
AirQuant	Airway morphometry	DL-based airway segmentation (U-Net)+rule-based tapering/tortuosity analysis	Airway diameter, tapering, tortuosity	[87,88]
AI-driven/data-driven
e-Lung	Automated fibrosis quantification	DL-based segmentation+fibrosis pattern quantification	%fibrosis burden	[89-91]
icolung	COVID/ILD lesion quantification; airway/vascular segmentation	DL-based lung/lobe/lesion segmentation	Lesion volume, %lung involvement, airway/vascular indices	[92-95]
SOFIA	Diagnostic classification of UIP/IPF	CNN-based image montage classifier → patient-level UIP probability	Probability scores for UIP (definite, probable, indeterminate, not UIP)	[96-98]
SATORI	Automated quantification of ILD extent	CNN-based segmentation of GGO, reticulation	%lung volume involved by GGO/reticulation	[99-101]
Qureight	Integrated fibrosis, airway, vascular, lung volume quantification	Multi-module segmentation+quantification+integration (Fib, Air, Vas, Lung)	%fibrosis, vascular metrics, airway morphometry, lung volume	[102,103]

QCT: quantitative computed tomography, ILD: interstitial lung disease, HU: Hounsfield unit, HAA: high-attenuation areas, LAA: low-attenuation areas, QLF: quantitative lung fibrosis, GGOs: ground-glass opacities, HC: honeycombing, FRI: functional respiratory imaging, CALIPER: Computer-Aided Lung Informatics for Pathology Evaluation and Ratings, VRS: vessel-related structures, ML: machine learning, DTA: data-driven textural analysis, QIA: quantitative interstitial abnormality, DL: deep learning, AI: artificial intelligence, COVID: Coronavirus disease, SOFIA: Systemic Objective Fibrotic Imaging Analysis Algorithm, UIP: usual interstitial pneumonia, IPF: idiopathic pulmonary fibrosis, CNN: convolutional neural network, SATORI: Segmentation and Annotation Tool for Radiomics and Deep Learning.

1) Densitometry

Densitometry, the most traditional and straightforward QCT approach based on CT density histograms, yields derived metrics such as the percentage of low-attenuation areas (%LAA, emphysema), high-attenuation areas (%HAA, ILD), mean lung attenuation (MLA), percentile density (PD15), and histogram shape indices including skewness and kurtosis. Densitometry-based methods, particularly the percentage of %HAA, have been applied to quantify ILAs. The concept of visually identified ILA was first proposed in the Genetic Epidemiology of COPD (COPDGene) cohort as a marker of subclinical CT changes associated with restrictive physiology (TLC <80% predicted; odds ratio [OR], 2.3; 95% CI, 1.4~3.7; p<0.001) [44]. Subsequently, a secondary analysis from the MESA study established %HAA as a quantitative imaging biomarker of ILA [45]. These concepts extended in patients with RA, secondary analysis from cardiac CT scans obtained in the Evaluation of Subclinical Cardiovascular Disease and Predictors of Events in RA (ESCAPE-RA) study demonstrated usefulness of %HAA in this population [46]. A high baseline %HAA, defined as >10% lung parenchymal involvement, was linked to female sex, greater smoking exposure (pack-years), elevated body mass index (BMI), and anti-cyclic citrullinated peptide (CCP) antibody titers ≥200 units. Over a median 3.5-year follow-up, nearly 50% of 106 participants demonstrated progressive increases in %HAA. This progression was associated with higher baseline %HAA, carriage of the MUC5B promoter variant, and absence of the HLA-DRB1 shared epitope. Notably, among patients with ≥5% increase in %HAA, higher baseline anti-CCP titers correlated with greater radiographic progression. Importantly, the clinical significance of %HAA appears particularly informative when serial measurements reveal worsening disease. A retrospective Canadian study analyzing 503 CT scans from 170 patients with SSc-ILD reported that each 1% increase in change of %HAA was associated with a 36% higher mortality risk, even after adjusting for age, sex, smoking exposure, baseline FVC, and DLCO [47].

2) Vitrea software

Vitrea software (Canon Group, Minnetonka, MN, USA) is a commercial densitometry platform and medical imaging tool designed for high-precision 3D visualization and volumetric quantification [48-50]. The software provides automated lung segmentation with options for manual correction, allowing the operator to obtain lung volume and density metrics, including low and HAA, with high precision.

In SARD-ILD, studies have investigated use of Vitrea software for the detection of PF in both SSc-ILD and RA-ILD [48,49], as well as the association between density-based grading and pulmonary function or prognosis [50]. In RA-ILD, lung quantification was performed using the density mask technique, which highlights voxels within specific density ranges: Low-Density Volume (−1,020 to −920 HU, associated with emphysema), Medium-Density Volume (−920 to −720 HU, representing functional lung tissue), and High-Density Volume (HDV, −720 to 0 HU, indicative of fibrosis). These quantitative assessments were compared against semi-quantitative methods such as the Goh scoring system and the Warrick method. The findings revealed significant associations between quantitative indices and the Goh score at baseline HRCT, with MLA and HDV also correlating with the Warrick extent score. Furthermore, during follow-up, all quantitative parameters correlated with both semi-quantitative measures. Despite these associations, neither quantitative nor semi-quantitative parameters at baseline predicted ILD progression.

3) Quantitative lung fibrosis/Quantitative interstitial lung disease

The most extensively investigated quantitative approach in SARD-ILD is the QLF and quantitative interstitial lung disease (QILD) assessment, initially developed in the context of SSc-ILD from SLS I, with the primary aim of quantifying PF [30,31]. QLF methodologies have focused on detecting reticular patterns and architectural distortion to estimate fibrotic burden. In SLS I, while the primary and secondary endpoints were PFTs and patient-reported outcomes including dyspnea and quality of life [37], application of QLF further revealed significant treatment effects: at 12 months, the cyclophosphamide arm demonstrated a decline in whole-lung QLF (−1.5%±9.5% vs. placebo +3.1%±5.7%, p=0.0190) and, more strikingly, in the zone of highest baseline QLF (−2.6%±20.8% vs. placebo +9.1%±12.6%, p=0.0027) [32]. Findings from SLS I and II also showed that QLF-based staging stratified patients as effectively as traditional visual scoring or the Goh–Wells criteria) [33].

The relationship between pulmonary function and quantitative imaging biomarkers has been well documented. Baseline QLF and QILD scores correlated most strongly with DLCO [35], whereas longitudinal changes in QLF and QILD were primarily associated with FVC decline over time [38]. Application of QILD has also facilitated a better understanding of treatment-related disease dynamics in SSc-ILD. In SLS, transition probabilities from fibrotic to GGO or normal patterns were significantly higher in the cyclophosphamide group, whereas the reverse transitions predominated in the placebo group [34].

Beyond SSc-ILD, QLF, and QILD have been applied to other autoimmune ILDs. In RA-ILD, QILD scores showed significant correlations with lung function: %DLCO with whole-lung (ρ=−0.433, p=0.027) and zone of maximal involvement (ρ=−0.501, p=0.009); %FVC with zone of maximal involvement (ρ=−0.395, p=0.046); and HRCT-derived TLC with QILD zone scores (ρ=0.637, p<0.001) [51]. Serum Krebs von den Lungen-6 (KL-6) levels also correlated with QILD (ρ=0.400, p=0.026) and QLF (ρ=0.566, p<0.001), but not with quantitative GGO (ρ=0.344, p=0.058) or HC scores (ρ=0.180, p=0.333). Another study demonstrated significant associations between QLF, QILD, and quantitative HC scores and visual assessment and further showed that these quantitative metrics independently predicted 5-year mortality after adjustment for age, erythrocyte sedimentation rate (ESR), FVC, and DLCO. A quantitative risk model combining age, ESR, and QLF scores effectively stratified mortality risk in RA-ILD [52].

In IIM-ILD, QILD scores also correlated strongly with lung function parameters. Whole-lung QILD was inversely correlated with FVC (r=−0.349, p=0.002) and DLCO (r=−0.381, p=0.001), while QILD of the most severely involved zone was correlated with FVC (r=−0.354, p=0.002) and DLCO (r=−0.340, p=0.003) [53]. Higher baseline QILD (>28.1%) and QLF (>10.4%) scores were associated with worse outcomes, as demonstrated by Kaplan–Meier survival curves (log-rank test: QILD, p=0.018; QLF, p=0.026). Furthermore, in antisynthetase syndrome-related ILD (ARS-ILD), QLF/QILD scores have shown both cross-sectional and longitudinal relevance. Analysis of 20 ARS-ILD patients from the longitudinal abatacept trial in myositis-associated ILD study showed that higher baseline whole-lung QLF and QILD were associated with lower baseline FVC and DLCO, and that longitudinal QILD trajectories were strongly associated with serial FVC, DLCO, dyspnea scores, Short Form-36 domains and oxygen requirements over 48 weeks [54]. More recently, the application of a growth-rate model to serial QLF scores in an IIM-ILD cohort (n=80) identified five distinct longitudinal patterns: progressive, improving, convex, others (mostly concave), and stable. This novel approach further revealed that patients exhibiting a rapid progression pattern (g>1.029%/month) had a significantly higher risk of mortality (HR, 15.926; 95% CI, 1.079~548.324; p=0.043) compared to the reference group, highlighting the prognostic utility of analyzing disease dynamics with QLF [55]. Representative images illustrating QLF/QILD are shown in Figure 1 [53].

Figure 1.

QLF/quantitative interstitial lung fibrosis. Representative serial images of a 57-year-old male patient with dermatomyositis-associated ILD showing concave-like dynamic QILD scores. (A~C) Axial HRCT images are displayed consecutively at baseline, 6 months and 26 months. (D~F) QILD scores are annotated on the axial images. Blue dots represent fibrosis, yellow dots represent GGOs, and pink dots represent honeycombs. (G~I) Coronal views of HRCT images and (J~L) QILD-annotated images are also shown. Each QILD score is indicated in the box below. QLF: quantitative lung fibrosis, ILD: interstitial lung disease, QILD: quantitative ILD, HRCT: high-resolution computed tomography, GGOs: ground-glass opacities, QGG: quantitative ground-glass opacity, QHC: quantitative honeycombing. Adapted from the article of Yeo et al. (Rheumatology (Oxford) 2023;62:3690-9) [53].

4) Functional respiratory imaging

Functional respiratory imaging (FRI) utilizes HRCT to reconstruct the three-dimensional architecture of the lungs, including airways, blood vessels, and lobes [56,57]. Building upon these reconstructions, computational fluid dynamics simulations are employed to quantitatively assess not only structural features but also functional metrics, such as airway volume, resistance, patterns of air distribution, and the deposition of inhaled medications within the pulmonary system. One notable advantage of FRI is their capacity to provide detailed anatomical and functional insights into the airways, in contrast to conventional PFTs. This precision makes such imaging particularly valuable for research on inhaled therapies, as it enables evaluation of regional drug delivery and efficacy at the airway level [58,59]. FRI received US Food and Drug Administration clearance for use in clinical practice received in March 2020 [K191550]. In SARD-ILD, there is one study that applied FRI to a limited number of patients with SSc-ILD [60]. In this study, positive correlation was observed between FVC and specific image-based airway resistance (siRaw; R²=0.07, p=0.046) and negative correlation between FVC and specific image-based airway radius (siRADaw; R²=0.24, p<0.001). At 12-months follow-up, patients in the moderate to severe group showed an increase from baseline in siRADaw (least square mean change 8.57% predicted; p=0.011) and decrease in siRaw (change –14.02% predicted; p=0.001). In patients with limited disease, there was no change in siRADaw (least square mean change 2.55% predicted; p=0.237) and a significant decrease in siRaw (change –25.11% predicted; p=0.006) from baseline.

5) Radiomics

Radiomics is a quantitative imaging methodology that systematically extracts and analyzes diverse features—including intensity, texture, morphology, and higher-order transformations—after segmentation of medical images, thereby linking tissue heterogeneity to clinical outcomes [61-63]. Unlike simple, first-order intensity measures such as the mean Hounsfield unit (HU) of lung CT, radiomics captures spatial heterogeneity in lung parenchyma, where density and patterns vary across regions. Texture features, by quantifying voxel intensity differences and spatial arrangements, reflect higher-order characteristics such as coarseness, cluster tendency, and sum of variance, rather than merely averaging voxel values. Wavelet transform–based radiomic features highlight subtle heterogeneity within lung tissue by extracting high-dimensional patterns. However, as these features rely on mathematically predefined transformations and formulas, they belong to the category of hand-crafted features. Consequently, similar to traditional quantitative CT approaches, their reproducibility requires validation across different wavelet decomposition schemes.

In the context of SARD-ILD, radiomics has been applied to SSc-ILD cohorts. In SSc-ILD, radiomics has been most extensively evaluated. Beyond the initial development of the quantitative radiomic risk score (qRISSc)—which stratified high- versus low-risk groups with HRs of 4.10 (95% CI, 1.87~9.03) in the derivation cohort (n=90) and 5.14 (95% CI, 1.14~23.20) in the validation cohort (n=66) [64]—a subsequent ERJ 2022 study expanded the radiomic approach by integrating radiomic texture clusters, pulmonary function parameters, and clinical variables into a multi-parameter nomogram to predict ILD progression in SSc [65]. In that study, radiomic features were among the strongest predictors of progression, and inclusion of radiomics significantly improved the C-index of the model compared with clinical and functional variables alone (radiomics-integrated model C-index 0.79 vs. clinical model 0.65). Radiomic clusters also reflected biologically relevant fibrotic pathways, indicating that radiomic markers capture lung-intrinsic, pathobiologically meaningful information that enhances prognostic stratification in SSc-ILD.

Beyond SSc-ILD, several recent studies have extended radiomics to RA-ILD and IIM-ILD. In RA-ILD, Venerito et al. [66] applied whole-lung HRCT radiomics in a proof-of-concept cohort and identified a four-feature Least Absolute Shrinkage and Selection Operator (LASSO)-selected signature in which higher median grey-level intensity strongly predicted mortality (HR, 9.35; 95% CI, 1.56~55.86), while the 10th percentile of included voxels, a voxel-based preprocessing index and flatness showed protective associations (HRs 0.20, 0.23, and 0.42, respectively) [67]. Han et al. [68] subsequently built a CT radiomics–KL-6 nomogram to identify low-risk RA-ILD, achieving areas under the receiver-operating-characteristic curve of 0.948 in the training set and 0.923 in the internal test set, and outperforming ILD-GAP (ILD–Gender, Age, Physiology index) staging alone. In anti-Melanoma Differentiation–Associated gene 5 (MDA5)-positive dermatomyositis–associated ILD, He et al. [69] analyzed 165 patients from two centers and showed that a bilateral-lung radiomics risk score predicted rapidly progressive ILD with the area under the curves (AUCs) of 0.869 and 0.905 in internal and external validation, while an integrated clinico-radiologic–radiomics nomogram achieved AUCs of 0.882 and 0.916 and identified high-risk groups with substantially increased mortality (HR 4.117 in the development cohort and 7.515 in the external cohort).

6) Computer-Aided Lung Informatics for Pathology Evaluation and Ratings

CALIPER is a supervised machine learning algorithm that classifies each HRCT voxel into one of five parenchymal categories—normal lung, emphysema, GGO, reticulation, or HC—based on canonical histogram signatures derived from expert-annotated training data. Total ILD volume is calculated as the sum of ground-glass, reticular, and HC areas and expressed as a percentage of total lung parenchymal volume (%ILD) [70]. Among CALIPER-derived indices, scores of vessel-related structures (VRS), encompassing pulmonary arteries and veins (excluding hilar vessels) as well as connected tubular structures reflecting fibrosis-associated vessel distortion, are of particular interest. VRS represents a distinct quantitative parameter apart from traditional CALIPER ILD scores and may reflect vascular remodeling that independently predicts mortality in SARD-ILD. In a cohort of 203 patients with SARD-ILD, CALIPER-derived pulmonary vessel volume (i.e., VRS) emerged as an independent predictor of mortality across all SARD-ILD subtypes, outperforming traditional visual and functional indices (patient age, smoking history, and carbon monoxide transfer coefficient); substitution of pulmonary function variables with VRS into the ILD-GAP model significantly improved prognostic performance (Stratified CT-GAP model AUC 0.74 vs. ILD-GAP model AUC 0.72) [71]. Building upon the recognition of VRS as an independent prognostic marker in SARD-ILD, subsequent analyses focused on its specific application in several specific SARDs. In RA-ILD, where fibrosing patterns are particularly common, a VRS threshold of ≥4.4% lung volume independently predicted mortality beyond established visual staging systems such as the scleroderma system and Fleischner IPF-based classification [72].

In a longitudinal SSc-ILD study by Occhipinti et al. [73], CALIPER showed stronger correlations with pulmonary function than semi-quantitative visual scoring (e.g., reticulation vs. FVC r=–0.71), and although individual CALIPER pattern changes did not predict functional decline, changes in total lung volume (TLV) predicted composite FVC/DLCO worsening (AUC 0.74, p=0.03). In another SSc-ILD cohort, CALIPER-derived vascular metrics demonstrated clear structure–function relationships, with total vessel number correlating with %FVC (r=–0.52) and total and venous vessel volumes correlating with %DLCO (r=–0.56 and –0.58). Small-caliber vessel counts (<6 mm) were higher in patients with ILD extent >5% (p<0.01), and reticulation showed parallel associations with %FVC and %DLCO [74]. Among 71 SSc-ILD patients, baseline reticulation ≥1.41% and ILD extent ≥6.32% were independently associated with mortality (Exp[B], 2.70; 95% CI, 1.26~5.82), while an absolute FVC decline ≥5% over 24 months was the strongest predictor of death (Exp[B], 15.01; 95% CI, 1.90~118.5) [75].

Several studies have also applied CALIPER to IIM-ILD. In one cohort of 110 patients, the percentage of total interstitial abnormalities correlated significantly with decreased TLC (r=–0.397, p<0.001), DLCO (r=–0.513, p<0.001), and oxygen saturation both at rest and during exercise (p<0.001 for both) [76]. Additionally, in a prospective study of 98 IIM-ILD patients, %ILD and %VRS measured by CALIPER demonstrated strong inverse correlations with %DLCO (ρ=–0.590 and –0.549, respectively; both p<0.001), and were closely associated with Warrick visual score (%ILD: ρ=0.663; %VRS: ρ=0.637; both p<0.001), with higher values observed in patients positive for anti–aminoacyl-tRNA synthetase antibodies [77].

CALIPER-assisted quantitative CT also offers potential for treatment monitoring. In a prospective study of patients with SSc-ILD or IPF, Di Battista et al. [78] demonstrated that serial CALIPER measurements could serve as outcome measures for antifibrotic therapy. Compared with a historical untreated cohort, patients receiving nintedanib exhibited stabilization of normal lung volume, ILD extent, reticular opacities, and GGOs. Notably, HC and VRS scores continued to increase irrespective of treatment, warranting further investigation. Furthermore, the CALIPER-derived ground-glass score was a strong predictor of DLCO decline; with a cutoff of 4.5, the relative risk for ≥10% DLCO worsening at 12 months was 6.8 (95% CI, 1.6~29.2) [79]. An example of CALIPER is shown in Figure 2 [72].

Figure 2.

Example of CALIPER. Axial CT image color maps demonstrating CALIPER-derived VRS in RA-ILD (red). VRS represent pulmonary arteries and veins (excluding hilar vessels) and connected tubular structures, the latter primarily reflecting adjoining regions of fibrosis. (A~C) Axial sections in a 71-year-old female 30-pack-year ex-smoker with upper lobe emphysema and fibrosis visible in the lower lobes (VRS 2.1%). (D~F) Axial sections in a 62-year-old female never-smoker with upper lobe-predominant fibrosis (VRS 7.0%). Nonvascular region captures in the VRS signal are visible in the upper lobes (D) and adjacent to the right hemidiaphragm (F). CALIPER: Computer-Aided Lung Informatics for Pathology Evaluation and Ratings, VRS: vessel-related structures, RA: rheumatoid arthritis, ILD: interstitial lung disease. Adapted from the article of Jacob et al. (Eur Respir J 2019;53:1800869) [72].

7) Data-driven textural analysis

Data-driven textural analysis (DTA) partitions HRCT voxels into small regions, extracts texture statistics (e.g., run-length, co-occurrence), and applies unsupervised clustering to automatically classify parenchymal patterns such as GGO, reticulation, and HC [80]. Unlike traditional feature engineering, where features are manually designed using predefined statistical or image-processing techniques, unsupervised feature learning derives features directly from the underlying data distribution. In patients with connective tissue disease-ILD, lower DTA-fibrosis score associated with lower HRs for 5-year lung transplant or death compared to high DTA-fibrosis score [81]. In RA-ILD, longitudinal change in DTA fibrosis score was associated with changes in FVC and DLCO (FVC percent predicted rho ¼–0.46; DLCO percent predicted rho ¼–0.43; both, p<0.001), and Cox multivariable models showed that the baseline DTA fibrosis score (HR, 1.03; 95% CI, 1.02~1.05; p<0.001) and changes in DTA fibrosis score over time (HR, 1.04; 95% CI, 1.02~1.06; p=0.003) were associated with increased risk of mortality with adjusted values based on multivariable model including age, sex, BMI, conventional disease-modifying antirheumatic drugs, biologics, smoking history, baseline DTA score, and change in DTA at follow-up [82].

8) Quantitative interstitial abnormality

DTA successfully quantified fibrosis burden at baseline and follow-up CT scans, but its region-based clustering approach imposed inherent limitations. Building on this, quantitative interstitial abnormality (QIA) was developed as a voxel-wise, machine learning–based CT method that integrates local density histograms with spatial features to automatically classify and quantify interstitial abnormalities such as GGOs, reticulation, and HC [83]. In two large prospective cohorts (4,636 participants from the COPDGene cohort and 1,307 participants from the Pittsburgh Lung Screening Study [PLuSS] cohort), annual change in QIA was associated with annual declines in FVC (COPDGene: 8.5 mL/y; 95% CI, 4.7~12.4 mL/y; p<0.001; PLuSS: 9.5 mL/y; 95% CI, 3.7~15.4 mL/y; p=0.001) and 6-minute walk distance, and increased mortality (COPDGene: HR, 1.69; 95% CI, 1.34~2.12; p<0.001; PLuSS: HR, 1.28; 95% CI, 1.10~1.49; p=0.001) [84]. In RA, McDermott et al. [85] analyzed association between genetic risk and ILA and QIA features. According to the study, while the RA polygenic risk scores (PRS) were not associated with ILA or QIA features, the IPF PRS was associated with both ILA (OR, 1.88 per unit; 95% CI, 1.52~2.32) and QIA (adjusted β=+0.50% per unit; p=7.3×10^-8) [85]. In COPDGene participants without known ILD or bronchiectasis, RA was also associated with a higher QCT-defined interstitial burden (median interstitial percentage 7.0% vs. 4.8%; p<0.0001). Moreover, RA combined with high emphysema burden (>75th percentile by QCT) conferred >5-fold higher mortality compared with non-RA with lower emphysema (HR, 5.56; 95% CI, 2.71~11.38) and RA with lower emphysema (HR, 5.86; 95% CI, 3.75~9.13) after adjusting for age, sex, smoking status, pack-years, and BMI [86].

9) AirQuant

AirQuant is an open-source image analysis pipeline that automatically computes airway lumen diameter, intersegmental tapering, and segmental tortuosity using hand-crafted (rule-based) algorithms. It requires an airway segmentation mask as input; in the index study, a 2D dilated U-Net (deep learning) was combined with the Pulmonary Toolkit (rule-based) to generate the mask, although other segmentation tools can be used [87,88]. This study analyzed 90 patients with IPF revealed that intersegmental tapering and segmental tortuosity as a significant predictor for mortality in IPF cohort adjusting for age, sex, smoking status, antifibrotic treatment and UIP pattern as well as FVC (HR, 0.75; 95% CI, 0.66~0.85; p<0.001 for intersegmental tapering; HR, 1.74; 95% CI, 1.22~2.47; p=0.002 for segmental tortuosity) or DLCO (HR, 0.769; 95% CI, 0.670~0.882; p<0.001 for intersegmental tapering; HR, 1.818; 95% CI, 1.204~2.746; p=0.004 for segmental tortuosity) or traction bronchiectasis extent (HR, 0.812; 95% CI, 0.724~0.911; p<0.001 for intersegmental tapering; HR, 1.491; 95% CI, 1.059~2.100; p=0.022 for segmental tortuosity) [88].

10) eLung

eLung (Branomix) is a deep learning-based algorithm that employs convolutional neural network (CNN) techniques to identify reticulation and pulmonary vascular structures on CT images. The eLung system automatically calculates the reticulovascular score (RVS) as well as the weighted RVS (WRVS), the latter being weighted according to the Euclidean distance from the peripheral regions of each lung [89-91]. Using the Open Source Imaging Consortium (OSIC) data, eLung measurement of 5% WRVS at baseline and 3% change of WRVS over 12 months were valuable to predict transplant-free survival (HR, 4.82; 95% CI, 2.66~8.73; p<0.01 for ≥15% vs. <15% WRVS at baseline; HR, 6.30; 95% CI, 3.34~11.89; p<0.001 for ≥3% vs. <3% change of WRVS over 12 months) in patients with IPF [91]. The usefulness and applicability of eLung needs to be assessed in SARD-ILDs.

11) icolung

icolung is an AI-based medical imaging software developed by icometrix, a Belgian company in Leuven, which has also developed a brain MRI analysis tool [92-95]. The software automatically segments the lung and characterizes pathological lesions, enabling quantitative assessment of high-attenuation abnormalities such as consolidation, GGOs, and reticulation. It calculates both the overall lesion burden relative to TLV and the extent of involvement in each lobe. Like other quantitative approaches, it allows longitudinal tracking of disease progression by comparing serial CT scans.

icolung uses deep learning–based CNNs performing voxel-level segmentation. The software provides automated quantification of TLV and integrates these results with lobe-based analysis. Abnormalities are visualized in three dimensions on coronal reconstructions and displayed as overlay maps after lobar segmentation. Severity scoring is performed per lobe, ranging from 0 (minimal involvement <1%) to 5 (>75% involvement), with a global score summing across all five lobes (0~25).

Initially developed for assessment of COVID-19 pneumonia [92,93], icolung has been applied in SARD-ILDs [94,95]. In SSc-ILD, quantitative lesion burden (%TLV) demonstrated strong inverse correlations with pulmonary function indices, including FVC (r=–0.50, p<0.0001) and DLCO (r=–0.46, p<0.0001) [94]. Progressive fibrosing ILD was characterized by a significantly higher TLV-adjusted abnormality burden than non-progressive ILD. Longitudinal analyses in 55 patients confirmed that changes in total interstitial abnormalities correlated with parallel declines in %FVC (r=–0.40, p=0.0025) and %DLCO (r=–0.40, p=0.0045). In RA-ILD, similar patterns were observed [95]. The proportion of interstitial abnormalities relative to TLV correlated with FVC (r=–0.24, p<0.01) and DLCO (r=–0.27, p<0.01). Among 49 patients, progressive PF (n=34) was associated with significantly higher abnormality percentages (2.45 [0.58~5.47] vs. 0.34 [0.08~1.21]; p<0.01) and CT severity scores (3 [1~5] vs. 1 [0~2]; p=0.01). These findings reinforce the utility of automated quantitative CT analysis as a biomarker of ILD progression. An example of icolung is shown in Figure 3 [94].

Figure 3.

Example of icolung. Automated icolung analysis in patients with SSc-ILD. (A) icolung software output combining the overall automatized lung segmentation (TLV quantification) and the association with lobar abnormalities. 3D analysis of parenchymal lung abnormalities: coronal view illustrating automated lung segmentation and the visualization of abnormalities. The abnormalities are quantified using icolung following lobe segmentation and represented with a conventional coronal view. The severity score is based on five lung lobes scoring on a scale of 0 to 5, with 0 indicating no involvement (<1%); 1, less than 5% involvement; 2, 5%~25% involvement; 3, 26%~49% involvement; 4, 50%~75% involvement; and 5, more than 75% involvement. The total severity score is the sum of the individual lobar scores and range from 0 (no involvement) to 25 (maximum involvement). (B) Screenshots of the icolung analysis report from baseline (left) and follow-up scan (right) of an anti–Scl-70–positive female patient with SSc on mycophenolate mofetil therapy exhibiting a progressive pulmonary fibrosis phenotype. FVC and DLCO (in % predicted) were 103% and 60% at baseline versus 60% and 28% at follow-up respectively. The 3D segmentation masks of the abnormalities are visualized in 2D axial and coronal views (red, consolidation; yellow, GGOs). SSc: systemic sclerosis, ILD: interstitial lung disease, TLV: total lung volume, ant-SCl-70: anti-topoisomerase I, FVC: forced vital capacity, DLCO: diffusing capacity for carbon monoxide, GGOs: ground-glass opacities, CNN: convolutional neural networ. Adapted from the article of Guiot et al. (Respir Res 2025;26:39) [94].

12) Systemic Objective Fibrotic Imaging Analysis Algorithm

Despite existing diagnostic criteria and guidelines for identifying and classifying fibrotic lung disease, their diagnosis can be challenging given the variability of distribution of fibrosis does not coincide with predominant pattern. The Systemic Objective Fibrotic Imaging Analysis Algorithm (SOFIA) is validated deep CNN tool that calculates the probabilities of UIP based on HRCT [96]. Using the Australian IPF registry (n=515), SOFIA based probability categories (SOFIA-the Prospective Investigation of Pulmonary Embolism Diagnosis, SOFIA-PIOPED) effectively categorized patients with their mortality risk and predicted transplant-free survival after adjusting for age, sex, and total ILD extent (HR, 1.29; 95% CI, 1.17~1.41; p<0.0001) [97]. Subgroup analysis with fibrotic ILD other than UIP or probable UIP by radiologists’ diagnosis as well as surgical lung biopsy proven possible, probable, or definite UIP (n=173) also revealed SOFIA-PIOPED UIP probability categories as valid predictors of transplant-free survival (HR, 1.34; 95% CI, 1.12~1.61; p<0.0001) along with total ILD extent (per 1% increments, HR, 1.02; 95% CI, 1.01~1.04; p<0.0001). In SSc-ILD, 27% (141/522) of patients exhibited intermediate/high UIP probability by SOFIA-PIOPED UIP probability [98]. Furthermore, those patients with intermediate/high UIP probability showed worse survival compared to patients with UIP not included in the differential (i.e., without UIP) by SOFIA-PIOPED in analyses adjusted for age, gender, ethnicity, smoking history, and either baseline FVC (HR, 1.88; 95% CI, 1.30~1.71; p=0.001) or Goh staging (HR, 1.33; 95% CI, 1.16~2.58; p=0.007).

13) Segmentation and Annotation Tool for Radiomics and Deep Learning

SATORI (Segmentation and Annotation Tool for Radiomics and Deep Learning) is an AI-based quantification of pulmonary HRCT, using the browser-based platform. SATORI was developed based on RACOON (Radiological COOperative Network), a multicenter radiological research network initiated during the COVID pandemic from the radiology departments of 38 German university hospitals [99]. SATORI performs automated quantification of lung volume (volumetry), GGO, reticulations, high attenuation lung volume (a histogram-based measurement of lung fibrosis via HUs) and emphysema. The extent of fibrosis using SATORI negatively correlated with FVC (ρ=–0.501), TLC (ρ=–0.622), and DLCO (ρ=–0.693) (p<0.001), and GGO also significant negatively correlated with DLCO (ρ=–0.699), TLC (ρ=–0.580), and FVC (ρ=–0.423) in patients with SARD-ILD [100]. A subsequent study assessing quantitative HRCT features in symptomatic and asymptomatic SARD-ILD patients reported that ILD involvement in asymptomatic patients was 8.9%±3.9% compared to 14.8%±11.0% in symptomatic patients (p=0.021) [101].

14) Qureight

Qureight consists of four deep learning models for quantifying lung volume (Lung8), vessel volume (Vascul8), fibrosis extent (Fibr8) and airway volume (Air8). Those specific CT-based biomarkers (the four imaging segmentations) were developed based on data from over 400 IPF patients in the Prospective Observation of Fibrosis in the Lung Clinical Endpoints (PROFILE) study and from a Cambridge cohort of IPF patients [102]. In the IPF patients, Lung8 showed strong correlation with FVC (r=0.82, p<0.001), vascular volume (HR, 1.30 [95% CI, 1.12~1.51]; p=0.001), and fibrosis volume (HR, 1.17 [95% CI, 1.12~1.22]; p=0.001) as well as Lung volume (HR, 0.98 [95% CI, 0.96~0.99]; p=0.001) were associated with 5-year survival adjusted for baseline GAP index. In analyzing 282 patients of the Australian IPF Registry (AIPFR), both SOFIA UIP probability (HR, 1.76; 95% CI, 1.12~2.76; p=0.01) and fibrosis volume (HR, 1.03; 95% CI, 1.02~1.04; p<0.001) were independently associated with transplant-free survival in multivariable analysis [103]. This method has not been applied to SARD-ILDs yet.

DISCUSSION

The purpose of imaging biomarkers is to enable early diagnosis, provide accurate prognostic assessment from baseline imaging, and facilitate longitudinal monitoring of therapeutic response. To achieve these objectives, considerable efforts have been directed toward the quantification of imaging data.

Qualitative and semi-quantitative CT assessments can be incorporated into routine clinical practice when guided by experienced thoracic radiologists. However, these methods rely heavily on subjective human interpretation and individual expertise, making them both time-consuming and prone to substantial inter-observer variability [40,41]. To address these limitations, computer-assisted quantitative approaches have been developed to improve precision and reproducibility. While such computational methods represent a significant advance, most proposed quantitative CT frameworks have yet to achieve widespread clinical adoption, primarily due to challenges with accessibility and reproducibility. Traditional approaches have relied on feature engineering, in which radiologists manually define and annotate imaging characteristics for computational measurement. This strategy is labor-intensive, dependent on specialist expertise, and inherently subjective. Moreover, because it is constrained by predefined assumptions about which CT features reflect disease severity or treatment response, it risks overlooking subtle but clinically important imaging findings. Deep learning offers a promising alternative by autonomously extracting informative features without the need for manual annotation or predefined rules. Through advanced pattern recognition, neural networks can emphasize clinically relevant signals in CT scans while minimizing irrelevant variation, thereby reducing reliance on resource-intensive human input. Importantly, this paradigm also enables the discovery of novel imaging biomarkers that are not detectable by conventional human observation. Nevertheless, the opacity of neural networks—the so-called “black box” problem—remains a major challenge, raising concerns about the reliability and reproducibility of their outputs. For deep learning methods to gain broader clinical acceptance, advancements in model interpretability and visualization, together with rigorous validation of biological plausibility, will be essential.

Methodologically, the QCT techniques reviewed here span a continuum from established feature-engineering paradigms (densitometry, Vitrea, QLF/QILD, radiomics, FRI, CALIPER, DTA, QIA, AirQuant) to novel deep learning architectures (eLung, icolung, SOFIA, SATORI, Qureight). Feature-engineered pipelines have demonstrated prognostic validity in well-phenotyped SSc-ILD and RA-ILD cohorts, occasionally being integrated into composite indices. Deep learning methods utilize learned representations derived from large datasets; however, these training sets are often dominated by IPF or mixed fibrotic ILDs, with specific application to SSc-ILD being a relatively recent extension. The SARD-ILD studies discussed in this review are summarized in Supplementary Table 1. Notably, there remain very limited data directly comparing multiple quantitative CT tools within the same SARD-ILD cohort; because most conclusions are inferred from separate studies, future research must prioritize head-to-head evaluations of these distinct computational approaches within harmonized multicenter cohorts.

While methodological advances are critical, the progress of quantitative CT also fundamentally depends on access to high-quality data. Developing and validating robust algorithms requires very large-scale imaging datasets; however, most currently available resources are retrospective, raising concerns about confounding by indication since CT scans are often performed during episodes of clinical deterioration. Collaborative efforts such as the OSIC (www.osicild.org) are encouraging, yet retrospective collections remain insufficient to fully address these limitations. To establish imaging biomarkers as reliable clinical endpoints, prospective cohorts that systematically integrate imaging, curated clinical data, and pulmonary function measures are indispensable. Moreover, explicitly incorporating imaging-derived outcomes as predefined endpoints in clinical trials may further strengthen their utility and accelerate their integration into clinical practice [104].

Furthermore, to achieve clinical applicability, quantitative CT methodologies should be integrated into routine imaging workflows. All quantitative CT methods are inherently dependent on HRCT acquisition. Traditional HRCT protocols, which obtain thin slices at fixed intervals and risk missing small pulmonary nodules. As a result, HRCT is unsuitable for lung cancer screening, where continuous volumetric acquisition with low-dose CT (LDCT) is required. Because patients with SARD-ILD face an elevated risk of lung cancer, imaging protocols should ideally be designed to address both disease assessment and cancer screening simultaneously. Recent advances in multidetector CT (MDCT) now allow continuous thin-section volumetric imaging that matches HRCT in resolution while offering full-lung coverage. Unlike interval-based HRCT protocols, this approach minimizes the likelihood of missing small pulmonary nodules, thereby enhancing diagnostic sensitivity. For this reason, MDCT is increasingly replacing conventional HRCT in routine practice. However, the higher radiation exposure of standard-dose MDCT (5~8 mSv) compared with LDCT (1~2 mSv, ~20% to 25% of a diagnostic chest CT dose) limits its utility for screening purposes. Fortunately, modern MDCT scanners can implement LDCT protocols while still producing thin-section volumetric datasets, thereby enabling a single scan to serve both lung cancer screening and quantitative CT analysis. Although increased noise and reduced signal-to-noise ratio in LDCT may affect the stability of certain features, particularly in radiomics and densitometry, advances in preprocessing and AI-based denoising techniques now allow both conventional QCT and deep learning–based methods to be applied reliably to LDCT images. Further validation studies are warranted to confirm that QCT approaches can be effectively applied to MDCT with LDCT protocols, achieving the dual goals of disease assessment and cancer screening in a single scan.

CONCLUSION

Quantitative imaging biomarkers such as QLF/QILD, CALIPER-derived vascular metrics, radiomics signatures, and deep-learning–based segmentation are reshaping the evaluation of SARD-associated ILD by providing objective, reproducible, and prognostically meaningful measures of parenchymal and vascular involvement. Across multiple cohorts, these approaches have shown consistent associations with lung function, disease progression, and, in selected settings, treatment response, suggesting that they can complement conventional physiologic and visual assessment rather than replace it.

However, their broader clinical adoption is limited by heterogeneous CT acquisition protocols, algorithm-specific variability, limited biological interpretability, and a relative paucity of prospective, multicenter validation. Moving forward, harmonized imaging standards, transparent and reproducible analytical pipelines, and clearly defined clinical thresholds will be crucial. Multidisciplinary efforts that integrate quantitative imaging with clinical, serological, and molecular data are needed to develop robust, externally validated prediction tools tailored to distinct SARD-ILD phenotypes. If these steps are achieved, quantitative CT and AI-derived biomarkers are likely to evolve from exploratory research instruments into integral components of routine care and clinical trial design, enabling earlier identification of high-risk patients, more precise therapeutic decision-making, and ultimately better outcomes for individuals living with SARD-ILD.

Further reading

For readers seeking detailed discussion on specific methodologies referenced in this review, we recommend the following literature:

• Visual scoring systems

º Assayag D, Kaduri S, Hudson M, Hirsch A, Baron M. High-resolution computed tomography scoring systems for evaluating interstitial lung disease in systemic sclerosis patients. Rheumatology (Sunnyvale) 2012;S1:003.

• Quantitative CT methods & reviews

º Kirby M, Smith BM. Quantitative CT scan imaging of the airways for diagnosis and management of lung disease. Chest 2023;164:1150-8.

º de la Orden Kett Morais SR, Felder FN, Walsh SLF. From pixels to prognosis: unlocking the potential of deep learning in fibrotic lung disease imaging analysis. Br J Radiol 2024;97:1517-25.

º Walsh SLF, De Backer J, Prosch H, Langs G, Calandriello L, Cottin V, et al. Towards the adoption of quantitative computed tomography in the management of interstitial lung disease. Eur Respir Rev 2024;33:230055.

º Dixon G, Thould H, Wells M, Tsaneva-Atanasova K, Scotton CJ, Gibbons MA, et al. A systematic review of the role of quantitative CT in the prognostication and disease monitoring of interstitial lung disease. Eur Respir Rev 2025;34:240194.

SUPPLEMENTARY DATA

Supplementary data can be found with this article online at https://doi.org/10.4078/jrd.2025.0161