Explainable Cryobiopsy AI Model, CRAI, to Predict Disease Progression for Transbronchial Lung Cryobiopsies with Interstitial Pneumonia

Explainable Cryobiopsy AI Model, CRAI, to Predict Disease Progression for Transbronchial Lung Cryobiopsies with Interstitial Pneumonia | medRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-P4HH5NV'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search Explainable Cryobiopsy AI Model, CRAI, to Predict Disease Progression for Transbronchial Lung Cryobiopsies with Interstitial Pneumonia View ORCID Profile Wataru Uegami , View ORCID Profile Ethan N. Okoshi , View ORCID Profile Kris Lami , Yuichiro Nei , Mutsumi Ozasa , Kensuke Kataoka , Yuka Kitamura , Yasuo Kohashi , Lee A.D. Cooper , View ORCID Profile Hidenori Sakanashi , Yuji Saito , Yasuhiro Kondoh , the study group on CRYOSOLUTION , View ORCID Profile Junya Fukuoka doi: https://doi.org/10.1101/2025.08.06.25333181 Wataru Uegami 1 Department of Pathology Informatics, Nagasaki University Graduate School of Biomedical Sciences , Nagasaki, Japan 2 Department of Pathology, Kameda Medical Center , Kamogawa, Japan Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Wataru Uegami Ethan N. Okoshi 1 Department of Pathology Informatics, Nagasaki University Graduate School of Biomedical Sciences , Nagasaki, Japan Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Ethan N. Okoshi Kris Lami 1 Department of Pathology Informatics, Nagasaki University Graduate School of Biomedical Sciences , Nagasaki, Japan Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Kris Lami Yuichiro Nei 3 Department of Respiratory Medicine, Respiratory Center , Toranomon Hospital, Minato-ku, Tokyo, Japan Find this author on Google Scholar Find this author on PubMed Search for this author on this site Mutsumi Ozasa 1 Department of Pathology Informatics, Nagasaki University Graduate School of Biomedical Sciences , Nagasaki, Japan Find this author on Google Scholar Find this author on PubMed Search for this author on this site Kensuke Kataoka 4 Department of Respiratory Medicine and Allergy , Tosei General Hospital, Seto, Japan Find this author on Google Scholar Find this author on PubMed Search for this author on this site Yuka Kitamura 1 Department of Pathology Informatics, Nagasaki University Graduate School of Biomedical Sciences , Nagasaki, Japan Find this author on Google Scholar Find this author on PubMed Search for this author on this site Yasuo Kohashi 5 Department of Respiratory Medicine, HARUHI Respiratory Medical Hospital , Kiyosu, Aichi, Japan Find this author on Google Scholar Find this author on PubMed Search for this author on this site Lee A.D. Cooper 6 Department of Pathology, Northwestern University Feinberg School of Medicine , Chicago, IL, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site Hidenori Sakanashi 7 Artificial Intelligence Research Center, National Institute of Advanced Industrial Science and Technology , Tsukuba, Ibaraki, Japan Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Hidenori Sakanashi Yuji Saito 5 Department of Respiratory Medicine, HARUHI Respiratory Medical Hospital , Kiyosu, Aichi, Japan Find this author on Google Scholar Find this author on PubMed Search for this author on this site Yasuhiro Kondoh 4 Department of Respiratory Medicine and Allergy , Tosei General Hospital, Seto, Japan Find this author on Google Scholar Find this author on PubMed Search for this author on this site Junya Fukuoka 1 Department of Pathology Informatics, Nagasaki University Graduate School of Biomedical Sciences , Nagasaki, Japan 2 Department of Pathology, Kameda Medical Center , Kamogawa, Japan Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Junya Fukuoka For correspondence: fukuokaj{at}nagasaki-u.ac.jp Abstract Full Text Info/History Metrics Supplementary material Data/Code Preview PDF Abstract Background Interstitial lung disease (ILD) encompasses diverse pulmonary disorders with varied prognoses. Current pathological diagnoses suffer from inter-observer variability,necessitating more standardized approaches. We developed an ensemble model AI for cryobiopsy, CRAI, an artificial intelligence model to analyze transbronchial lung cryobiopsy (TBLC) specimens and predict patient outcomes. Methods We developed an explainable AI model, CRAI, to analyze TBLC. CRAI comprises seven modules for detecting histological features, generating 19 pathologically significant findings. A downstream XGBoost classifier was developed to predict disease progression using these findings. The model’s performance was evaluated using respiratory function changes and survival analysis in cross-validation and external test cohorts. Findings In the internal cross-validation (135 cases), the model predicted 105 cases without disease progression and 30 with disease progression. The annual Δ%FVC was −1.293 in the non-progressive group versus −5.198 in the progressive group, outperforming most pathologists’ diagnoses. In the external test cohort (48 cases), the model predicted 38 non-progressive and 10 progressive cases. Survival analysis demonstrated significantly shorter survival times in the progressive group (p=0.034). Interpretation CRAI provides a comprehensive, interpretable approach to analyzing TBLC specimens, offering potential for standardizing ILD diagnosis and predicting disease progression. The model could facilitate early identification of progressive cases and guide personalized therapeutic interventions. Funding New Energy and Industrial Technology Development Organization (NEDO) and Japanese Ministry of Health, Labor, and Welfare. 1 Introduction Interstitial lung disease (ILD) is a diverse group of pulmonary disorders that vary in prognosis and treatment approaches. 1 For instance, idiopathic pulmonary fibrosis (IPF) is a progressive condition with a poor prognosis, often treated by antifibrotic medications. 2 In contrast, non-specific interstitial pneumonia (NSIP), hypersensitivity pneumonitis 3 (HP), and connective tissue disease-associated ILDs 4 (CTD-ILDs) typically require treatment targeting inflammatory changes or the underlying causes, such as inhaled antigens or autoimmune conditions. Therefore, accurate diagnosis is crucial for determining appropriate treatment strategies and predicting patient outcomes. Recent advancements have revealed the presence of poor prognostic subgroups within non-IPF interstitial lung diseases. 1 , 5 The latest IPF guideline introduced the concept of progressive pulmonary fibrosis (PPF), which is defined as at least two of the following criteria: worsening respiratory symptoms, physiological evidence of disease progression, and radiological evidence of disease progression occurring within the past year in patients with ILDs other than IPF. 1 Additionally, multiple clinical trials have tested antifibrotic drugs as a treatment for various progressive ILDs. 6 – 8 There is growing attention on methods to identify progressive cases early to perform therapeutic interventions at the appropriate time. Pathological examinations are promising in that they can be conducted at an early stage of diagnosis and provide a wealth of information; however, histological studies on progressive cases are still few and were not included in the 2022 criteria for PPF. Tsushima et al. have reported that the focal usual interstitial pneumonia (UIP) pattern is a key prognostic pattern in PPF. 9 UIP is characterized by temporally and spatially heterogeneous fibrosis, manifesting as a patchy distribution of immature and dense fibrotic areas. 1 , 10 This pattern reflects the progressive nature of the disease and has significant implications for patient prognosis and treatment decisions. However, the reliability of pathological evaluations has been challenged by low inter-observer agreement among pathologists. 11 , 12 This inconsistency raises significant concerns, as it can lead to misdiagnoses, resulting in patients receiving inappropriate treatments or missing out on beneficial therapies, which may accelerate disease progression. The advent of artificial intelligence (AI)-based image analysis offers promising potential for standardizing pathological diagnoses and addressing the challenge of inter-observer variability. 13 , 14 Our previous work introduced MIXTURE, the first and only reliable AI model designed to analyze surgical lung biopsy (SLB) specimens in interstitial lung disease. 15 MIXTURE demonstrated high accuracy in detecting UIP patterns in SLB samples, with patients classified as UIP by the model showing significantly poorer survival outcomes. While SLB has traditionally been the gold standard for the pathological diagnosis of ILD, its invasive nature 16 has limited its application. In contrast, the less invasive transbronchial lung cryobiopsy (TBLC) has gained prominence. 1 , 17 However, SLB and TBLC specimens differ substantially in both qualitative and quantitative aspects, presenting unique challenges for AI analysis. In this study, we present CRAI for cryobiopsy, a novel ensemble AI model specifically designed for TBLC specimens. This model incorporates seven generic models that evaluate hematoxylin and eosin (H&E) specimens to highlight comprehensive histopathological features in a pathologically interpretable manner. These models’ findings are fed into a downstream model to predict disease progression and patients’ prognosis. By providing a more accurate and consistent analysis of TBLC specimens, our model has the potential to improve ILD treatment by offering critical guidance for physicians’ treatment decisions. 2 Materials and Methods Patients and Datasets Four datasets were constructed from the two institutions ( Table 1 ). The TGH-SLB set was collected at Tosei General Hospital and consulted at Nagasaki University, containing pathological whole slide images (H&E and EVG staining) of SLB cases. This dataset was used for model training to supplement the limited number of TBLC cases. The TGH-TBLC set, used for model training and validation, consisted of TBLC cases collected at Tosei General Hospital and consulted at Nagasaki University. This dataset included histopathological slides, pulmonary function test (PFT) data (% forced vital capacity, %FVC), and survival data. The baseline PFT was the most recent test performed less than 100 days prior to the TBLC procedure, and follow-up PFTs at one and two years were also collected when available. The HRH-PFT and HRH-survival set, used as a held-out external test set, included consecutive TBLC cases diagnosed at Haruhi Respiratory Medical Hospital (HRH) for PFT analysis and survival analysis respectively. View this table: View inline View popup Download powerpoint Table 1: Summary of cohort characteristics and baseline details. Data are n, n (%) or mean±SD. PFT=pulmonary function test, SLB=surgical lung biopsy, TBLC=transbronchial lung cryobiopsy, TBLB=transbronchial lung biopsy, FVC=forced vital capacity, D L CO =Diffusing capacity for carbon monoxide, prog=progressioin. Disease progression was defined as ≥5% absolute decline in percentage for each year over a two-year observation period. All pathological specimens were scanned using an Aperio CS2 whole slide image (WSI) scanner at 20× or 40× magnification. This research plan was approved by the Ethics Committee of Nagasaki University Hospital (approval number: 19081929-4). Model Overview CRAI consists of seven different modules that detect basic histological structures. By integrating these structures at the slide level, 17 pathologically significant slide-level features are calculated. Finally, a downstream XGBoost classifier is constructed using these 17 findings as input to predict patients’ disease progression ( Figure 1 ). The models are summarized in Table S1 and the examples of each models’ outputs are presented in Figures S1 and S2. Download figure Open in new tab Figure 1: Workflow of CRAI: A Multi-model Approach for Prognostic Prediction. Trans-bronchial lung cryobiopsy specimens undergo comprehensive analysis through seven specialized deep learning models: the lung-stroma model, classification models optimized for 1×, 5×, and 20× magnifications, and models for collagen, elastica, and lymphocyte detection. Each model extracts distinct pathological features, which are then consolidated to generate 17 pathologically interpretable slide-level findings. These parameters are subsequently integrated into an XGBoost-based downstream model for disease progression prediction. Created in https://BioRender.com Lung-Stroma Model This model is designed to distinguish the bronchovascular bundle and lung parenchymal regions, which are relatively abundant in TBLC specimens. The model was implemented as a segmentation model for bronchovascular bundles in the specimen. Annotations were provided for 45 TBLC whole slide images, and an EfficientNet-b0 18 backbone U-Net 19 was trained on 512px (3mm) square patches. 1× Mode006C This model is designed to detect the distribution of three distinct tissue patterns: “honeycomb”, “patchy”, and “diffuse”, which can be recognized at low magnification (1.25×). A total of 192 slides (72 cases) diagnosed with UIP from the TGH-SLB dataset were utilized as training data. The patch size was set to 460 pixels (3.68mm) square. A pathologist specializing in ILD classified the patches as “patchy” using a web application (available at https://github.com/uegamiw/sortIT ). Patches classified as “patchy” were defined as those containing a mixture of lesions and normal lung tissue. In a similar manner, patches identified as “honeycomb” were selected from 72 slides (57 cases) diagnosed with UIP, where the honeycomb pattern is characterized by cystic spaces surrounded by dense fibrosis. Conversely, patches considered to be “diffuse” were selected from 192 slides (71 cases) diagnosed as Non-UIP, with the diffuse pattern defined as patches where lesions occupy the area without any normal lung tissue. This classification resulted in 2,268, 1,617, and 1,704 unique patches for the honeycomb, patchy, and diffuse categories, respectively. Common features of the patches were extracted using KimiaNet, 20 and a multi-layer perceptron (MLP) regressor head was trained to calculate the probability of each feature’s presence. Optimal hyperparameters were determined through grid search with 5-fold cross-validation. 5× and 20× Models These multi-class classification models classify tissue patterns at a moderate (5×) and high magnification (20×). It utilizes MIXTURE, a previously published model design from our earlier work. 15 To avoid semantic overlap with the newly developed model, some classes were merged as shown in Table S2 and S3. Collagen and Elastic Fiber Models These models quantitatively estimate the deposition of two distinctive fibers, collagen and elastic fibers, from H&E images in cases of ILD. For the TGH-SLB set, registration of H&E and EVG specimens was performed using Deep Feature Based Registration. 21 The implementation was modified from the TIA toolkit code. 22 The registration results were reviewed, and 37 out of 150 successful WSIs were selected. A pathologist performed color deconvolution on these 37 EVG specimens using QuPath, 23 separating collagen fibers (red) from elastic fibers (black). The localization information of these fibers was saved as 1/8 size PNG mask images. Based on this, corresponding patches of H&E specimens of the same size as the 20× model were matched with patches of EVG specimens and the resulting color deconvolution outputs for collagen and elastica masks. The deposition levels of collagen and elastic fibers in the H&E images were defined as the ratio of the mask areas of the corresponding collagen and elastica masks, and a regressor model was constructed to infer the deposition levels of each fiber. The model architecture was ResNet18 pre-trained on ImageNet. Lymphocyte Model This model detects lymphocytes. Inflammatory cells were identified using HoverNet 24 with weights from the PanNuke dataset 25 resulting in 194,881 segmentation instances. A YOLOv7-based detection model was then constructed for rapid identification of inflammatory cells in 20× images. Slide Level Integration From the features identified by each model and their combinations, the following 18 findings were obtained: (1) lung, (2) lesion, (3) fibrotic/cellular ratio, (4) lymphoid follicles, (5) dense fibrosis, (6) bronchial epithelium, (7) number of fibroblastic proliferation/organizing pneumonia (FP/OP), (8) maximum collagen around the FP/OP, (9) lymphocyte density in dense fibrosis area, (10) lymphocyte density in non-dense fibrosis, (11) collagen in non-dense fibrosis area, (12) collagen edge, (13) collagen edge/dense fibrosis, (14) collagen forming the blob, (15) elastic fiber, (16) diffuse, (17) patchy, (18) honeycomb. The definition for each finding is shown in the supplementary materials. Cases where the lung area was estimated to be less than 10 mm 2 were excluded due to insufficient evaluable regions. All features other than lung were adopted as slide-level findings. Construction of the Downstream Model for the Disease Progression Prediction From the TGH-TBLC cohort, 135 cases with follow-up respiratory function data over 2 years were extracted. Progressive disease was defined as a decline of 5% or more in %FVC or a decline of 10% or more in percentage of the diffusing capacity of the lungs for carbon monoxide (% DL CO ) within any single year, accompanied by no improvement in both functional tests from the baseline after two years of follow-up. The remaining cases were classified as the non-progressive group and served as labels for the training data; these served as labels for the training data. Using the 17 features extracted by AI as input, an XGBoost classifier was constructed to classify prognosis. XGBoost was selected based on our previous methodological experience. 26 Hyperparameters were optimized through ten-fold cross-validation and grid search. Thus, for the model with the best hyperparameters, ten XGBoost models were trained, one for each cross-validation split. For the inference values of the TGH-TBLC cohort, the inference results were synthesized from the cross-validation results. For external validation using the HRH-PFT cohort, the mode of the outputs from the above ten models was adopted as the output. UIP Diagnosis by Multiple Pathologists 28 pathologtists independently observed 135 TBLC specimens, the same as above (TGH-TBLC). Referring to the WSIs, they classified the UIP pattern, which is well known as a histopathological prognostic factor. UIP patterns were categorized into two groups: UIP (present throughout the sample or occupying the majority of the lesion) and non-UIP (occupying a minor area of the lesion or completely absent). Analysis was performed using data from the 19 pathologists who diagnosed 130 or more cases. Evaluation of CRAI and Pathologists’ UIP to Predict Δ%FVC Patients with multiple recorded PFT measurements were selected for the study. The absolute change in %FVC (Δ%FVC) is considered an important predictor of mortality. 27 For each pathologist, we applied a linear mixed model which included fixed effects for age, sex, and baseline %FVC, and a random effect for the interaction between UIP diagnosis and time (in years) to evaluate the differences in Δ%FVC by diagnosis group. To compare the pathologists’ diagnoses with CRAI, the same analysis was conducted using the output of CRAI (non-progressive vs. progressive group) instead of the UIP diagnosis. Statistical significance was determined using a 95% confidence interval. Survival Analysis This analysis was performed on the HRH-survival cohort. Overall survival was defined as the time from the date of clinical diagnosis to the date of death from any cause or the end of follow-up. Survival estimates were calculated using the Kaplan–Meier method for patient subgroups classified by AI, and these estimates were compared using the log-rank test. The level of significance was established α =0.05. Subset Analysis for Interstitial Lung Abnormality (ILA) Test cases from TGH included cases categorized as interstitial lung abnormality based on clinical criteria. 28 A subset analysis was conducted separately to evaluate prediction of progression in these cases. Environment The deep learning environment used Python 3.8.19 and Torch 1.13.1. Scikit-learn 1.3.0 and xgboost 2.1.1 was used for constructing the XGBoost models, and statsmodels 0.14.1 was used for the generalized linear mixed model analysis. Survival analysis was performed in the R 4.4.1 environment. The log-rank test used the survival package 3.6, and the survminer package 0.5.0 was used for plotting the Kaplan-Meier curves. 3 Results Predictive Performance for Respiratory Function Deterioration In the cross-validation of the TGH-TBLC cohort, CRAI predicted 105 cases as a non-progressive group and 30 cases as a progressive group. In the non-progressive group, the estimated annual Δ%FVC was −1.293, whereas in the progressive group, it was −5.198, with a statistically significant difference between the two groups ( Figure 2 , Table S4). Thirty patients were clinically categorized as having ILA, with ten of them identified as having a disease progression. The estimated Δ%FVC exhibited a similar trend (−2.853 in the non-progressive group and −8.093 in the progressive group, Table S5). Among the cases of ILD, some examples of the cases that were classified as progressive group by the model and actually showed clinically worsening respiratory function are shown in Figure S3. In external validation using the HRH-PFT cohort, 21 cases were inferred as a non-progressive and nine cases a progressive. The Δ%FVC in the non-progressive and progressive groups was −1.926 and −5.894, respectively, closely reflecting the results of the cross-validation cohort (Table S6). Download figure Open in new tab Figure 2: Comparison of prognostic performance between CRAI and pathologists using Generalized Linear Mixed Models. The forest plot demonstrates the estimated annual change in %FVC for different diagnostic categories in the TGH-TBLC cohort. The topmost row shows CRAI’s predictions, which classifies each case as progressive or non-progressive. The subsequent rows display the predictions based on 19 pathologists’ diagnoses, comparing UIP versus non-UIP classifications. The x-axis represents the estimated annual change in %FVC, with negative values indicating decline. Points and horizontal lines represent coefficient estimates and their 95% confidence intervals, respectively. CRAI demonstrated separation between non-progressive versus progressive groups without overlapping confidence intervals, indicating statistical significance. Among the 19 pathologists, only five achieved non-overlapping confidence intervals between their UIP and non-UIP diagnoses. These results suggest that CRAI’s prognostic performance surpasses that of most pathologists in predicting functional decline. The model was adjusted for age, sex, and baseline %FVC. FVC=forced vital capacity, UIP=usual interstitial pneumonia. In the pathological diagnosis of interstitial pneumonia, the histopathological indicator for predicting patient survival prognosis is the presence or absence of UIP. Therefore, we compared the predictions of CRAI with each pathologist’s UIP diagnosis from the perspective of annual Δ%FVC. In the TGH-TBLC cohort, only five out of 19 pathologists could differentiate respiratory function prognosis ( Figure 2 ). The model’s performance exceeded the prognostic prediction accuracy of many pathologists specializing in interstitial pneumonia. The interobserver agreement of UIP judgement among 19 pathologists were considered as low (Fleiss K = 0.293). Comparison of UIP Diagnosis and CRAI We analyzed the prognosis predictions of CRAI and the diagnostic tendencies of pathologists for UIP. Although there were significant differences in UIP diagnostic tendencies between pathologists, it was found that most pathologists’ UIP diagnoses were similar to those of the model. We investigated the distribution of the cases using UMAP analysis based on whole slide level features. The cases classified as progressive group by CRAI corresponded with the cases diagnosed as UIP by many pathologists (Figure S4). Predictive Performance for Survival Prognosis In the HRH-survival cohort, the median duration of follow-up was 1,561 days. At the last follow-up, 15 (31.2%) patients had died. The model predicted 38 cases as non-progressive and 10 cases as a progressive group, out of a total of 48 cases. In the non-progressive group, seven patients died, whereas in the progressive group, eight patients died. The group classified as progressive disease had a statistically significantly shorter survival time compared to the group inferred as a non-progressive group ( Figure 3 ) (p=0.034). Download figure Open in new tab Figure 3: Kaplan-Meier Survival Curves for Progression Classification in the External Cohort. The progressive group as predicted by CRAI showed significantly shorter survival times compared to the non-progressive group (log-rank test, p=0.038). The risk table below the graph shows the number of patients at risk in each group over time. These results demonstrate the model’s ability to effectively stratify patients based on their prognosis in an independent cohort from a different institution. Survival analysis was conducted on the HRH-survival cohort. Pathological Characteristics of Cases Classified as Progressive Group by CRAI SHAP (SHapley Additive exPlanations) analysis indicated that cases classified as progressive by the model showed a correlation between the presence of elastic fibers or nodular collagen deposition and the progressive group. Additionally, a correlation was observed between lymphoid follicles or FP/OP and the non-progressive group (Figure S5). 4 Discussion We have developed an explainable AI models, CRAI, the first AI model specifically designed to predict disease progression from TBLC specimens. The model offers comprehensive pathological analysis, and the results suggest its utility as a prognostic predictor. Notably, the model achieves considerable performance in differentiating pulmonary function deterioration. This result is particularly impressive considering that prognosis-predictive diagnosis of UIP has been challenging for most pathologists, even those experienced in ILD diagnosis. Low agreement of UIP diagnosis is confirmed by the low kappa coefficient, K =0.293 by the 19 pathologists in the TGH-TBLC cohort. The results suggest that, in many cases, the model could outperform prediction of deteriorating respiratory function compared to histological UIP diagnosis given by expert pulmonary pathologists. The definition of short-term or long-term prognosis for ILD remains controversial. While the criteria for PPF are well-accepted, the setting of the observation period and the threshold for assessing respiratory function are arbitrary, and depending on the observation term, cases are categorized differently as seen in our TGH cases ( Table 1 ). In this context, respiratory function serves as a well-known and independent surrogate marker for patient survival. 29 , 30 Due to the ease of data acquisition, we adopted it as the primary feature for establishing ground truth labels in the training data, while also utilizing it to compare the AI model’s and pathologists’ diagnoses. However, given the significant variability in PFT results, we also prepared survival time as a more robust outcome for an external cohort. Consequently, in the external cohort, there was no significant difference in predicted %FVC between the progressive and non-progressive groups, but the estimated annual %FVC changes closely resembled the cross-validation results. In the survival analysis of the external validation cohort, the group predicted as a progressive group exhibited significantly shorter survival times, supporting the model’s robustness. The cases predicted by CRAI as a progressive group generally possessed the features recognized by pathologists as UIP, aligning with the known association between UIP and progressive disease. 9 , 12 Interestingly, SHAP analysis indicated that nodular collagen fiber deposition and elastosis were progression factors, while fibroplasia (FP/OP) and lymphoid follicles were factors associated with non-progressive group, demonstrating similarity with the diagnostic process of UIP by pathologists. Instead of adopting an end-to-end model using recent large foundation models that directly predicts outcomes such as respiratory function prognosis or survival prognosis, CRAI emphasized extracting pathologically interpretable findings. This approach is particularly suitable in the context of ILD diagnosis, which differs qualitatively from tumor diagnosis. While tumors exhibit significant variations in size, shape, and surrounding microenvironments, ILD presents as a mixture of inflammatory and post-inflammatory reactions composed of normal cellular components. As a result, the requirement for specific and granular morphological parameter recognition is less critical for ILD AI models compared to cancer models. It is also important for cases requiring explainability and for those with a limited amount of digital data compared to common neoplastic diseases. While pathologists can easily recognize findings such as fibrosis or inflammation, the lack of standardized quantification methods for these features has led to variability in final diagnoses. These detected findings are not only utilized for the specific downstream task such as the patient’s prognosis but also serve as a platform for various histopathological analyses. Compared to our previous MIXTURE model targeting SLB, 15 this model incorporates several improvements considering the histological characteristics of TBLC. One key difference is the sampling site. In SLB, the peripheral parts of the lung are sampled, and large bronchovascular bundles are less likely to be included. In contrast, TBLC specimens are collected transbronchially, inevitably including bronchi and surrounding tissues as a major part of biopsy sample. Since diagnostically important lesions are primarily located in the lung parenchyma rather than in the bronchovascular bundles, and SLB samples are large enough to minimize the influence of bronchovascular structures, however; TBLC tissues are comparatively small, making the presence of bronchovascular bundles more impactful within the same tissue. Therefore, separating these areas from evaluation in the AI model is crucial. This study has limitations. A major issue is the small number of cases available for consideration to cover disease progression. TBLC is a relatively recently introduced technique, and in some regions, only a limited number of institutes can perform. In particular, it is challenging to collect cases with long-term follow-up necessary to evaluate true disease progressive behavior with such a new technique. However, in places such as Japan and Europe, TBLC is increasingly being adopted by more facilities due to its less invasive nature. As more cases accumulate in the future, further improvement in the model’s performance and validation will become possible. The current model focuses solely on predicting prognosis from the histological findings of TBLC. With this approach, it may be possible to identify potentially progressive cases without observing them over a certain period, as per the criteria for PPF. It could detect progressive cases at the stage of ILA as our data indicated, which are often incidentally discovered in asymptomatic individuals during screening. 28 Furthermore, by combining multiple modalities such as clinical data, physiological measurements, and radiological imaging, it could enable more precise identification of progressive patients than is currently possible, leading to more personalized and timely therapeutic interventions as our previous work. 31 In summary, we have developed CRAI, the first deep learning-based model to comprehensively analyze ILD TBLC tissues. This model enables stratification of disease progression and will not only provide a foundation for objective tissue analysis of ILD but also lead to standardized diagnosis to optimize patient care. 5 Data Sharing Upon approval from the Institutional Review Board at Nagasaki University Hospital, a subset of individual participant data can be shared for research purposes upon reasonable request to the corresponding author. Our experiments utilized the open-source code of YoLoV7 ( https://github.com/WongKinYiu/yolov7 ). The code for the CRAI system is not available for sharing, as it is currently undergoing a business transfer process. Data Availability Upon approval from the Institutional Review Board at Nagasaki University Hospital, a subset of individual participant data can be shared for research purposes upon reasonable request to the corresponding author. 6 Acknowledgement This paper is based on results obtained from a project, JPNP20006, commissioned by the New Energy and Industrial Technology Development Organization (NEDO) and partly supported by the research group of diffuse lung disease under the Japanese Ministry of Health, Labor, and Welfare (20FC1033). Footnotes ↵ † CRYOSOLUTION Wendy Cooper (Tissue Pathology and Diagnostic Oncology, NSW Health Pathology, Royal Prince Alfred Hospital), Luka Brcic (Department of Pathology, Hospital Graz II, Graz, Department of Pathology, Medical University of Vienna), Yoshiaki Zaizen (Nagasaki University), Joanne Lynne Wright (Department of Pathology, St Paul’s Hospital, University of British Columbia), Angela Maria Takano Pena (Singapore General Hospital and Duke-NUS Medical School), Sosuke Ishijima (Nagasaki University Graduate School of Biomedical Sciences), Amna R. Almutrafi (Toronto General Hospital, Research Institute), Mary Beth Beasley (Mount Sinai Medical Center), Alberto Cavazza (Azienda USL Reggio Emilia), Izidor Kern (University Clinic Golnik), Wei-Chin Chang (Taipei Medical University Hospital), Lee Fong Wan (Sarawak General Hospital), Tereza Losmanova (Institute of Pathology, University of Bern), Noriah binti Othman (Hospital Serdang), Alexandre Todorovic Fabro (Ribeirão Preto Medical School – University of São Paulo), Masayo Yoshimura (Fukuoka University School of Medicine and Hospital), Akira Hebisawa (Asahi General Hospital), Koji Okudera (Saitama Medical University), Rosane Duarte Achcar (National Jewish Health), Andrew Churg (The University of British Columbia, Vancouver General Hospital), and Richard Luther Attanoos (Cardiff University) References [1]. ↵ Raghu G , Remy-Jardin M , Richeldi L , Thomson CC , Inoue Y , Johkoh T , et al. Idiopathic pulmonary fibrosis (an update) and progressive pulmonary fibrosis in adults: an official ATS/ERS/JRS/ALAT clinical practice guideline . Am J Respir Crit Care Med . 2022 ; 205 ( 9 ): e18 – 47 . OpenUrl CrossRef PubMed [2]. ↵ King Jr TE , Bradford WZ , Castro-Bernardini S , Fagan EA , Glaspole I , Glassberg MK , et al. A phase 3 trial of pirfenidone in patients with idiopathic pulmonary fibrosis . N Engl J Med . 2014 ; 370 ( 22 ): 2083 – 92 . OpenUrl CrossRef PubMed Web of Science [3]. ↵ Vasakova M , Selman M , Morell F , Sterclova M , Molina-Molina M , Raghu G. Hypersensitivity pneumonitis: current concepts of pathogenesis and potential targets for treatment . Am J Respir Crit Care Med . 2019 ; 200 ( 3 ): 301 – 8 . OpenUrl CrossRef PubMed [4]. ↵ Mathai SC , Danoff SK . Management of interstitial lung disease associated with connective tissue disease . BMJ . 2016 ; 352 . [5]. ↵ Nasser M , Larrieu S , Boussel L , Si-Mohamed S , Bazin F , Marque S , et al. Estimates of epidemiology, mortality and disease burden associated with progressive fibrosing interstitial lung disease in France (the PROGRESS study) . Respir Res . 2021 ; 22 ( 1 ): 162 . OpenUrl CrossRef PubMed [6]. ↵ Flaherty KR , Wells AU , Cottin V , Devaraj A , Walsh SL , Inoue Y , et al. Nintedanib in progressive fibrosing interstitial lung diseases . N Engl J Med . 2019 ; 381 ( 18 ): 1718 – 27 . OpenUrl CrossRef PubMed [7]. Behr J , Neuser P , Prasse A , Kreuter M , Rabe K , Schade-Brittinger C , et al. Exploring efficacy and safety of oral Pirfenidone for progressive, non-IPF lung fibrosis (RELIEF)-a randomized, double-blind, placebo-controlled, parallel group, multi-center, phase II trial . BMC Pulm Med . 2017 ; 17 : 1 – 9 . OpenUrl CrossRef PubMed [8]. ↵ Maher TM , Corte TJ , Fischer A , Kreuter M , Lederer DJ , Molina-Molina M , et al. Pirfenidone in patients with unclassifiable progressive fibrosing interstitial lung disease: a double-blind, randomised, placebo-controlled, phase 2 trial . Lancet Respir Med . 2020 ; 8 ( 2 ): 147 – 57 . OpenUrl CrossRef PubMed [9]. ↵ Tsushima Y , Okoshi EN , Ishijima S , Bychkov A , Lami K , Morimoto S , et al. Presence of focal usual interstitial pneumonia is a key prognostic factor in progressive pulmonary fibrosis . Histopathology . 2024 ; 85 ( 1 ): 104 – 15 . OpenUrl PubMed [10]. ↵ Lynch DA , Sverzellati N , Travis WD , Brown KK , Colby TV , Galvin JR , et al. Diagnostic criteria for idiopathic pulmonary fibrosis: a Fleischner Society White Paper . Lancet Respir Med . 2018 ; 6 ( 2 ): 138 – 53 . OpenUrl CrossRef PubMed [11]. ↵ Mäkelä K , Hodgson U , Piilonen A , Kelloniemi K , Bloigu R , Sutinen E , et al. Analysis of the histologic features associated with interobserver variation in idiopathic pulmonary fibrosis . Am J Surg Pathol . 2018 ; 42 ( 5 ): 672 – 8 . OpenUrl PubMed [12]. ↵ Hashisako M , Tanaka T , Terasaki Y , Uekusa T , Achcar RD , Aswad BI , et al. Interobserver agreement of usual interstitial pneumonia diagnosis correlated with patient outcome . Arch Pathol Lab Med . 2016 ; 140 ( 12 ): 1375 – 82 . OpenUrl PubMed [13]. ↵ Lami K , Bychkov A , Matsumoto K , Attanoos R , Berezowska S , Brcic L , et al. Overcoming the interobserver variability in lung adenocarcinoma subtyping: a clustering approach to establish a ground truth for downstream applications . Arch Pathol Lab Med . 2023 ; 147 ( 8 ): 885 – 95 . OpenUrl CrossRef PubMed [14]. ↵ Bulten W , Balkenhol M , Belinga JJA , Brilhante A , Ç akır A , Egevad L , et al. Artificial intelligence assistance significantly improves Gleason grading of prostate biopsies by pathologists . Mod Pathol . 2021 ; 34 ( 3 ): 660 – 71 . OpenUrl CrossRef PubMed [15]. ↵ Uegami W , Bychkov A , Ozasa M , Uehara K , Kataoka K , Johkoh T , et al. MIXTURE of human expertise and deep learning—developing an explainable model for predicting pathological diagnosis and survival in patients with interstitial lung disease . Mod Pathol . 2022 ; 35 ( 8 ): 1083 – 91 . OpenUrl CrossRef PubMed [16]. ↵ Han Q , Luo Q , Xie JX , Wu LL , Liao LY , Zhang XX , et al. Diagnostic yield and postoperative mortality associated with surgical lung biopsy for evaluation of interstitial lung diseases: a systematic review and meta-analysis . J Thorac Cardiovasc Surg . 2015 ; 149 ( 5 ): 1394 – 401 . OpenUrl CrossRef PubMed [17]. ↵ Troy LK , Grainge C , Corte TJ , Williamson JP , Vallely MP , Cooper WA , et al. Diagnostic accuracy of transbronchial lung cryobiopsy for interstitial lung disease diagnosis (COLDICE): a prospective, comparative study . Lancet Respir Med . 2020 ; 8 ( 2 ): 171 – 81 . OpenUrl PubMed [18]. ↵ Tan M , Le Q. Efficientnet: Rethinking model scaling for convolutional neural networks ; 2019 . [19]. ↵ Ronneberger O , Fischer P , Brox T. U-net: Convolutional networks for biomedical image segmentation ; 2015 . [20]. ↵ Riasatian A , Babaie M , Maleki D , Kalra S , Valipour M , Hemati S , et al. Fine-tuning and training of densenet for histopathology image representation using tcga diagnostic slides . Med Image Anal . 2021 ; 70 : 102032 . OpenUrl CrossRef PubMed [21]. ↵ Awan R , Raza SEA , Lotz J , Weiss N , Rajpoot N. Deep feature based cross-slide registration . Elsevier ; 2023 . [22]. ↵ Pocock J , Graham S , Vu QD , Jahanifar M , Deshpande S , Hadjigeorghiou G , et al. TIAToolbox as an end-to-end library for advanced tissue image analytics . Commun Med . 2022 ; 2 ( 1 ): 120 . OpenUrl PubMed [23]. ↵ Bankhead P , Loughrey MB , Fernández JA , Dombrowski Y , McArt DG , Dunne PD , et al. QuPath: Open source software for digital pathology image analysis . Sci Rep . 2017 ; 7 ( 1 ): 1 – 7 . OpenUrl CrossRef PubMed [24]. ↵ Graham S , Vu QD , Raza SEA , Azam A , Tsang YW , Kwak JT , et al. Hover-net: Simultaneous segmentation and classification of nuclei in multi-tissue histology images . Med Image Anal . 2019 ; 58 : 101563 . OpenUrl CrossRef PubMed [25]. ↵ Gamper J , Koohbanani NA , Benes K , Graham S , Jahanifar M , Khurram SA , et al. Pannuke dataset extension, insights and baselines ; 2020 . [26]. ↵ Sano H , Okoshi EN , Tachibana Y , Tanaka T , Lami K , Uegami W , et al. Machine-Learning-Based Classification Model to Address Diagnostic Challenges in Transbronchial Lung Biopsy . Cancers . 2024 ; 16 ( 4 ). [27]. ↵ Du Bois RM , Weycker D , Albera C , Bradford WZ , Costabel U , Kartashov A , et al. Forced vital capacity in patients with idiopathic pulmonary fibrosis: test properties and minimal clinically important difference . Am J Respir Crit Care Med . 2011 ; 184 ( 12 ): 1382 – 9 . OpenUrl CrossRef PubMed Web of Science [28]. ↵ Ichikado K , Ichiyasu H , Iyonaga K , Kawamura K , Higashi N , Johkoh T , et al. An observational cohort study of interstitial lung abnormalities (ILAs) in a large Japanese health screening population (Kumamoto ILA study in Japan: KILA-J) . BMC Pulm Med . 2023 ; 23 ( 1 ): 199 . OpenUrl PubMed [29]. ↵ Gimenez A , Storrer K , Kuranishi L , Soares MR , Ferreira RG , Pereira CA . Change in FVC and survival in chronic fibrotic hypersensitivity pneumonitis . Thorax . 2018 ; 73 ( 4 ): 391 – 2 . OpenUrl Abstract / FREE Full Text [30]. ↵ Jegal Y , Kim DS , Shim TS , Lim CM , Do Lee S , Koh Y , et al. Physiology is a stronger predictor of survival than pathology in fibrotic interstitial pneumonia . Am J Respir Crit Care Med . 2005 ; 171 ( 6 ): 639 – 44 . OpenUrl CrossRef PubMed Web of Science [31]. ↵ Lami K , Ozasa M , Che X , Uegami W , Kato Y , Zaizen Y , et al. Enhancing interstitial lung disease diagnoses through multimodal AI integration of histopathological and CT image data . Respirology . 2025 . View the discussion thread. Back to top Previous Next Posted August 08, 2025. Download PDF Supplementary Material Data/Code Email Thank you for your interest in spreading the word about medRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following Explainable Cryobiopsy AI Model, CRAI, to Predict Disease Progression for Transbronchial Lung Cryobiopsies with Interstitial Pneumonia Message Subject (Your Name) has forwarded a page to you from medRxiv Message Body (Your Name) thought you would like to see this page from the medRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share Explainable Cryobiopsy AI Model, CRAI, to Predict Disease Progression for Transbronchial Lung Cryobiopsies with Interstitial Pneumonia Wataru Uegami , Ethan N. Okoshi , Kris Lami , Yuichiro Nei , Mutsumi Ozasa , Kensuke Kataoka , Yuka Kitamura , Yasuo Kohashi , Lee A.D. Cooper , Hidenori Sakanashi , Yuji Saito , Yasuhiro Kondoh , the study group on CRYOSOLUTION , Junya Fukuoka medRxiv 2025.08.06.25333181; doi: https://doi.org/10.1101/2025.08.06.25333181 Share This Article: Copy Citation Tools Explainable Cryobiopsy AI Model, CRAI, to Predict Disease Progression for Transbronchial Lung Cryobiopsies with Interstitial Pneumonia Wataru Uegami , Ethan N. Okoshi , Kris Lami , Yuichiro Nei , Mutsumi Ozasa , Kensuke Kataoka , Yuka Kitamura , Yasuo Kohashi , Lee A.D. Cooper , Hidenori Sakanashi , Yuji Saito , Yasuhiro Kondoh , the study group on CRYOSOLUTION , Junya Fukuoka medRxiv 2025.08.06.25333181; doi: https://doi.org/10.1101/2025.08.06.25333181 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Pathology Subject Areas All Articles Addiction Medicine (568) Allergy and Immunology (863) Anesthesia (300) Cardiovascular Medicine (4435) Dentistry and Oral Medicine (444) Dermatology (382) Emergency Medicine (608) Endocrinology (including Diabetes Mellitus and Metabolic Disease) (1509) Epidemiology (15227) Forensic Medicine (30) Gastroenterology (1124) Genetic and Genomic Medicine (6597) Geriatric Medicine (668) Health Economics (997) Health Informatics (4534) Health Policy (1368) Health Systems and Quality Improvement (1613) Hematology (540) HIV/AIDS (1264) Infectious Diseases (except HIV/AIDS) (15916) Intensive Care and Critical Care Medicine (1103) Medical Education (623) Medical Ethics (146) Nephrology (667) Neurology (6599) Nursing (346) Nutrition (998) Obstetrics and Gynecology (1144) Occupational and Environmental Health (957) Oncology (3332) Ophthalmology (974) Orthopedics (369) Otolaryngology (420) Pain Medicine (436) Palliative Medicine (130) Pathology (663) Pediatrics (1693) Pharmacology and Therapeutics (691) Primary Care Research (711) Psychiatry and Clinical Psychology (5447) Public and Global Health (9230) Radiology and Imaging (2198) Rehabilitation Medicine and Physical Therapy (1370) Respiratory Medicine (1196) Rheumatology (593) Sexual and Reproductive Health (712) Sports Medicine (530) Surgery (712) Toxicology (99) Transplantation (289) Urology (265) (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'a004052f9830ad07',t:'MTc3OTUzODg4NQ=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();