Automated AI-based image analysis for quantification and prediction of interstitial lung disease in systemic sclerosis patients

Abstract

Background

Systemic sclerosis (SSc) is a rare connective tissue disease associated with rapidly evolving interstitial lung disease (ILD), driving its mortality. Specific imaging-based biomarkers associated with the evolution of lung disease are needed to help predict and quantify ILD.

Methods

We evaluated the potential of an automated ILD quantification system (icolung^®) from chest CT scans, to help in quantification and prediction of ILD progression in SSc-ILD. We used a retrospective cohort of 75 SSc-ILD patients to evaluate the potential of the AI-based quantification tool and to correlate image-based quantification with pulmonary function tests and their evolution over time.

Results

We evaluated a group of 75 patients suffering from SSc-ILD, either limited or diffuse, of whom 30 presented progressive pulmonary fibrosis (PPF). The patients presenting PPF exhibited more extensive lesions (in % of total lung volume (TLV)) based on image analysis than those without PPF: 3.93 (0.36–8.12)* vs. 0.59 (0.09–3.53) respectively, whereas pulmonary functional test showed a reduction in Force Vital Capacity (FVC)(pred%) in patients with PPF compared to the others : 77 ± 20% vs. 87 ± 19% (p < 0.05). Modifications of FVC and diffusing capacity of the lungs for carbon monoxide (DLCO) over time were correlated with longitudinal radiological ILD modifications (r=-0.40, p < 0.01; r=-0.40, p < 0.01 respectively).

Conclusion

AI-based automatic quantification of lesions from chest-CT images in SSc-ILD is correlated with physiological parameters and can help in disease evaluation. Further clinical multicentric validation is necessary in order to confirm its potential in the prediction of patient’s outcome and in treatment management.

Background

Systemic sclerosis (SSc) is a rare connective-tissue disease of unknown origin affecting multiple organs. Characterized by autoimmunity, vessel inflammation, and organ fibrosis, SSc is classified by skin fibrosis extent into two patterns: limited cutaneous systemic sclerosis (lcSSc) and diffuse cutaneous systemic sclerosis (dcSSc). A significant complication of SSc is interstitial lung disease (ILD), which varies from slow to rapid progression [1,2,3,4,5]. Its clinical classification is based on the extent of skin fibrosis, which divides patients into two major patterns: limited cutaneous systemic sclerosis (lcSSc), which is characterised by skin fibrosis limited to the elbows and knees; and diffuse cutaneous systemic sclerosis, which involves proximal areas, the face, and the trunk in addition to distal areas [6, 7]. A main complication of SSc that contributes to morbidity is the appearance of interstitial lung disease (ILD) [8, 9]. The clinical course of SSc-associated interstitial lung disease (SSc-ILD) can range from a slowly progressing lung disease to a rapid progression [1]. The challenge, as with other ILDs [4], is identifying patients at high risk of progression and initiating early therapeutic intervention including immunosuppressive and/or antifibrotic therapy in order to limit disease extension and clinical flare up [2].

Biomarkers are crucial for quantifying ILD in SSc, providing insights into disease activity, severity, progression, and treatment response. They aid in early detection, risk stratification, distinguishing ILD subtypes, and guiding targeted treatment. Radiological ILD quantification on high-resolution CT (HRCT) is vital for monitoring disease progression, predicting outcomes, and identifying patients for early intervention [10]. However visual assessment of HRCT in ILD is prone to high inter-observer variability, poor reproducibility, and relative insensitivity to subtle disease progression over short follow-up periods. Machine learning models enhance HRCT assessments by accurately detecting and quantifying lung abnormalities, offering reproducible evaluations and reducing inter-observer variability. These models also provide objective treatment response measures, improving drug development and patient selection in clinical trials [11,12,13,14,15,16,17,18,19,20].

In this study, we investigate the automatic AI-based quantification of interstitial lung abnormalities from HRCT in SSc-ILD patients and their correlations with standard clinical markers like pulmonary function tests.

Methods

Patients cohort

We retrospectively analyzed 75 patients with SSc-ILD from the ILD clinic of University Hospital of Liège (CHU) seen between 9th of january 2007 to 30th of september 2022. Demographic data were collected (age, gender, BMI, tobacco status). Concerning systemic sclerosis classification, patients were classified according to the 2013 ACR/EULAR criteria for SSc [6] and the distinction of the cutaneous forms into limited and diffuse was made according to the classification of Leroy et al. [7]. Sine scleroderma are patients without cutaneous disease with other organ involvement. Concerning ILD classification, patients were classified as progressive pulmonary fibrosis if they met INBUILD criteria (PF-ILD) which was the definition used before 2022 [21]. Criteria for fibrotic progression of ILD were, on a follow-up period of 24 months before analysis:

• a relative decline in the forced vital capacity (FVC) of at least 10% of the predicted value;

• or a relative decline in the FVC of 5% to less than 10% of the predicted value and worsening of respiratory symptoms or an increased extent of fibrosis on high-resolution computed tomography (HRCT) of the chest;

• or worsening of respiratory symptoms and an increased extent of fibrosis on HRCT.

A total of 55 patients had longitudinal follow-up CT scans (between 2 and 4 scans), among which 23 were PF-ILD patients. The protocol was approved by ethics committee of CHU Liège, (Belgian number: B7072020000033) and all experiments were performed in accordance with relevant guidelines and regulations.

Pulmonary function tests

Lung function tests were performed in the routine respiratory laboratory at CHU Liège in accordance with the recommendations of the European Respiratory Society (ERS) [22]. Volumes are expressed in liter (L)(absolute value) and as percentage of predicted normal values. The Tiffeneau index or FEV1/FVC is expressed as percentage. The diffusion capacity of CO (DLCO) and carbon monoxide transfer coefficient (KCO or DLCO/VA) were measured by the single breath testing technique (Sensor Medics 2400 He / CO Analyzer System, Bilthoven, The Netherlands).

Statistical analysis

Parametric distribution of continuous variables were described using means and standard deviations (SD) and non-parametric distributions were described using median (interquartile range: IQR) expressed as number (% Yes). Spearman correlation coefficient (r) [23] was used for correlation between PFT and imaging parameters (strong correlation r > 0.7; moderate 0.3–0.7; low < 0.3). Paired T-test and Wilcoxon matched-pairs signed rank test for parametric and non-parametric variables respectively [24] were used for the longitudinal analysis. The comparison between SSc-ILD and SSc-PFILD was analysed by unpaired T-test for FVC and DLCO and by Mann-Whitney test for TLV% and Severity score [25]. A p value less than 0.05 was accepted for statistical significance. Statistical analysis was performed using the TIBCO Statistica, v. 13.5.0, TIBCO Software Inc, Palo Alto, CA, USA and graphs using GraphPad Prism software version 9.0.0 for Windows, GraphPad Software, San Diego, California, USA software package.

Imaging parameters

All the CT images used in the study were acquired on one of the five multidetector CT scanners: Siemens Edge Plus (2), GE Revolution CT (1), and GE Brightspeed (2). Since CT images were collected retrospectively, no standardized scan protocol was available over the complete dataset. All scans were non-contrast High Resolution CT (slice thickness ≤ 1 mm) and acquired as per standard of care.

Icolung software

We employed the icolung software, which performs fully automatized segmentation of the lungs, lung lobes, and lung abnormalities (ground-glass opacity and consolidation) consecutively using deep learning models. These convolutional neural network models are based on the 2D and 3D U-net architectures described in [26, 27], and were trained, validated, and tested on clinical CT scans, along with voxel-level delineations of lung abnormalities, created by expert radiologists. Based on the models’ predicted masks for lung abnormalities and lobes, the lung involvement in each lobe was computed as the ratio of abnormality volume vs. lobe volume, from which was derived a lobe-specific severity score (0–5). The five severity scores were then summed into a global severity score (0–25) for the patient’s current CT exam. An example of the software report is depicted in Fig. 1. The software output consisted of Total Lung volume (TLV) (L), Lung abnormalities (combining consolidation and ground-glass opacities)(% of TLV), Consolidation (% of TLV), Ground-Glass Opacities (% TLV) and severity score per patient.

Fig. 1

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 a 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 SSc female patient, SCL-70 + treated with MMF exhibiting a PPF 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 = ground glass opacities)

Results

The baseline clinical, functional and imaging parameters of the patients are reported in Table 1. P values are reported for statistically significant differences between the SSc-ILD and SSc-PFILD groups.

Table 1 Patients’ clinical, functional characteristics and imaging-based quantifications

The correlation between the lesion percentage out of the total lung volume (Lesion %) and the results of the pulmonary function tests (FVC and DLCO) is reported in Fig. 2. Both lung FVC and DLCO had a significant correlation with the lesion extent percentage extracted from the image analysis (r=-0.50 and r=-0.46 respectively). The complete correlation analysis is reported in Table S1.

Fig. 2

Correlation between IA quantification of lesions and functional parameters. Correlation between percentage of lesions quantified out of the TLV from the AI algorithm with FVC (A) and DLCO (B)

The comparison between SSc-ILD and SSc-PFILD patients is reported in Fig. 3. Statistically significant differences were found between the two groups in terms of FVC (Fig. 3A), TLV% (Fig. 3B) and Severity score (Fig. 3D). Differences in DLCO values were not statistically significant (Fig. 3C).

Fig. 3

Comparison between functional and imaging biomarkers in SSc-ILD and SSc-PFILD. Difference in FVC (A), TLV% (B), DLCO (C) and Severity score (D) for patients with SSc-ILD (blue bar) and SSc-PFILD (green bar). Data was analyzed by unpaired T test for FVC and DLCO and by Mann Whitney test for TLV and Severity score. *p value < 0.05

Patients with SSc-PFILD displayed a higher extent of lesions out of the Total Lung Volume ( % TLV): 3.93 (0.36–8.12) versus 0.59 (0.09–3.53)(p < 0.05) and a higher Severity score: 4 (1–6) versus 1 (0–4) than those of the SSc-ILD group (p < 0.05). Pulmonary functional tests showed a marked decrease in FVC: 77 ± 20% pred. for SSc-PFILD in comparison to the other group (FVC: 87 ± 19% pred.)(p < 0.05). The complete results of the comparison tests are reported in Table S1 for the imaging parameters and PFTs.

A subsequent analysis was performed on 55 patients with longitudinal data (32 SSc-ILD and 23 SSc-PFILD), considering T1 as the scan at diagnosis and T2 as follow-up scan. Figure 4 reports the variation over time of FVC (Fig. 4A), DLCO (Fig. 4B) and TLV% (Fig. 4C). The variation over T1 and T2 of TLV% showed moderate correlation with both the variation of FVC (r= -0.40, p < 0.01) and DLCO (n = 48)(r= -0.40, p < 0.01) in Fig. 4D and E, respectively. The mean follow-up was 3.9 years (1.8–6.7)(p = 0.45). The complete correlation analysis is reported in Table S1.

Fig. 4

ILD quantification compared to PFT modifications over time. A.B.C. Variation over time of FVC (A), DLCO (B) and TLV % (C). Data was analyzed by the paired T test for FVC and DLCO and by the Wilcoxon matched-pairs signed rank test for TLV %. T1 and T2 represent the two different timepoints for CT-scan analysis *P value < 0.05 ;**P value < 0.01. D.E. Correlation between the variation over the time of FVC (D) and DLCO (n = 48)(E) correlated with ILD progression over time out of the total lung volume (TLV %)

Discussion

Quantification of interstitial lung disease on CT scans is still challenging for clinicians. Qualitative and semiquantitative visual assessments of disease extent in ILDs have been adopted, the latter striving for quantification and standardization of the evaluation of disease extent in comparison to the former. However, both approaches suffer from moderate to high inter-reader variability (kappa value ranging from 0.28 to 0.85) [28], necessitate description of the location of the abnormalities allowing a basic quantification of ILD severity [29,30,31].

The differentiation of the abnormalities and the grading of each abnormal area still remains a challenge. Several different abnormalities can be present in ILD patients and even the most common, namely ground glass opacities (GGO), consolidation, reticulation and honeycombing, are difficult to differentiate and correctly attribute [32]. Moreover, these qualitative and semi-quantitative approaches are time consuming, dependent on radiological expertise and prone to manual errors. In the present manuscript, we describe the application of an automated, reproducible, and accurate quantification tool that can help radiologists and clinicians overcome these limitations, allowing a fully quantitative approach.

Novel solutions to unburden clinical staff, to help in the diagnosis process and evaluation of changes over time represent an urgent clinical need. In the last years, several additional deep learning (DL) methods have been put out for the purpose of identifying and quantifying lung abnormalities [33, 34], which can be then used to ILD lesions [35, 36], highlighting the potential benefits of this approach in the management of patients with restrictive lung diseases. Furthermore, several studies have reported correlations between this automatic AI-based quantification of lung abnormalities and PFT parameters used in standard of care clinical practice for diagnosis and follow up. Si-Moahmed et al. [37] used a commercially available software tool to measure the lung CT volume and correlate this with FVC and total lung capacity (TLC). Sue et al. [38] explored the correlation between abnormalities pattern volume and vasculature volume with PFT parameters in a cohort of idiopathic pulmonary fibrosis (IPF) patients. In the specific scenario of SSc-related ILDs, Karadag et al. [39] investigated automatic extraction of textural lung features and their relationship to pulmonary function tests and visual fibrosis scores (VFS), finding moderate to strong correlations, and a notable distinction between PF and non-PF patients.

Our novel approach investigates the temporal progression of both progressive and non-progressive SSc-ILD patients, which is, to our knowledge, the first study to do so. It also clearly establishes a correlation between automatic volumetric quantification of abnormalities and PFTs, enabling the characterization and distinction of both groups of patients. The ability to follow this change over time is of the utmost importance for therapy planning and follow-up, and it might promote the development of innovative imaging-based endpoints in clinical trials when evaluating effectiveness of antifibrotic medications [40, 41]. Adding automatized HRCT quantification tools in patients follow-up is of major interest as a clinical decision support companion aiming to reliably help clinicians in patients management.

The present approach combining AI-based image analysis and AI, tackles three of the key challenges in the management of ILD patients: early diagnosis, accurate prognostication from baseline scans and therapy response monitoring over time [42]. The automatic quantification of lung abnormalities could in turn streamline and simplify current classification criteria for ILDs, distinguishing between limited and extensive disease in a quantitative, robust and reproducible way, for example with an improved Goh algorithm for optimal prognosis of SSc-ILD patients [43, 44]. Moreover, the correlations between both PFTs and imaging parameters in the cohort of progressive versus non progressive patients, may pave the road for the future development of an integrated and automated image analysis tool able to predict patients with higher risk of developing PPF [45, 46]. This will improve patients’ management and result in better quality of life, when therapies can be delivered faster and more efficiently in a context of personalized medicine.

The current study has some limitations. The data are monocentric and retrospectively collected, so there is no homogenous image acquisition protocol nor scan acquisition timeline. Moreover, we also included patients with limited lung involvement (< 10%) in the cohort. While this represents better the reality of clinical practice, the lack of homogeneous imaging parameters and comparable scan acquisition times might have reduced the performance and affect especially the longitudinal analysis and the overall correlation PFTs. We also used the PFILD definition as the patients were clinically classified before the publication of the guidelines. Nevertheless, the case review confirmed that all patients were also presenting the new PPF criteria. A prospective trial is envisioned in the future to provide more homogeneous data collecting and potentially extent longitudinal analysis to more than one follow-up scan. In addition, the current version of the software identifies and quantifies only ground glass opacities and consolidation. However, some reports in literature indicate variables correlations between other abnormal patterns and PFTs results, for example reticulation which seems to be correlated with DLCO (r = -0.581) [47, 48]. Taking into account other abnormalities patterns in future analysis might refine even more the diagnostic and prognostic approach based on automatic AI quantification from chest CT scans.

Conclusion

Quantitative metrics obtained from AI-driven analysis of chest CT images in SSc-ILD has shown promising results in the correlations with PFTs, supporting quantifiable and reproducible disease evaluation. This approach holds the potential to improve the management of SSc-ILD patients. However, prior to its integration into routine clinical practice, it is necessary to perform comprehensive clinical multicentric studies to validate the model outcomes and their correlation with standard of care PFTs. Along with this effort, we need to elucidate its ability to accurately predict patient outcomes, especially regarding the insurgence of progressive pulmonary fibrosis. The possibility of predicting PPF from baseline scans as well as response to therapy is an urgent clinical unmet need in the field of SSc-ILD.