TabPFN-driven ternary classification of stage IA lung adenocarcinoma subtypes using AI-derived histogram features a retrospective multicenter cohort study.

Introduction
Lung adenocarcinoma (LUAD), the most prevalent nonsmall-cell lung cancer subtype, remains a leading global cause of cancer mortality[1]. Stage IA lesions exhibit heterogeneous prognosis dictated by histological invasiveness per the 2021 WHO classification[2]. Precise preoperative differentiation among precursor glandular lesions (PGL), minimally invasive adenocarcinoma (MIA), and invasive adenocarcinoma (IAC) is critical for guiding personalized surgical strategies. Distinguishing PGL from MIA holds particular clinical relevance: PGL (e.g., atypical adenomatous hyperplasia) may warrant conservative sublobar resection or even active surveillance, whereas MIA – defined by ≤5 mm stromal invasion – requires complete resection with negative margins to minimize recurrence risk[2,3]. In contrast, binary classification (noninvasive vs. invasive) fails to stratify intermediate-risk lesions, potentially leading to overtreatment of indolent PGL or undertreatment of early-invasive MIA. While thin-section CT remains the cornerstone for nodule evaluation, conventional imaging criteria achieve only 58–65% accuracy in subtyping due to overlapping radiological features[4]. Radiomics and deep learning (DL) show promise in quantifying tumor heterogeneity through histogram analysis, yet most studies focus on binary classification rather than the clinically critical ternary distinction[5,6].
Significant translational challenges persist despite advances in AI-based pulmonary nodule analysis[7–9]. First, existing models for LUAD subtyping primarily address binary classification tasks (e.g., invasive vs. noninvasive) and lack robust multicenter validation[10]. Second, while histogram features derived from DL-driven segmentation may capture subtle tumor heterogeneity[11], their predictive value for LUAD subtyping remains underexplored. Third, traditional machine learning (ML) algorithms (e.g., logistic regression, random forests) struggle with high-dimensional, correlated feature spaces common in radiomics, leading to overfitting and suboptimal performance on multiclass tasks[12]. Moreover, current studies rarely benchmark model performance against intraoperative frozen section analysis – the clinical gold standard for intraoperative decision-making[13]. These limitations underscore the need for robust, interpretable models capable of leveraging multimodal data to preoperatively stratify LUAD subtypes across diverse populations.
To address these challenges, we developed a novel framework integrating Tabular Prior-data Fitted Network (TabPFN) – a transformer-based tabular foundation model – with AI-derived histogram features for ternary classification of histological subtypes[14]. Building upon validated AI systems for pulmonary nodule analysis[7–9], this multicenter study aims to: (1) develop and validate discriminative models for PGL/MIA/IAC using state-of-the-art algorithms; (2) objectively compare performance against both traditional machine learning methods and intraoperative frozen section analysis; and (3) identify potential patient subgroups deriving maximal benefit through stratified performance analysis. This work aligns with the 2021 WHO Classification of Lung Tumors, emphasizing the clinical necessity of preoperative LUAD subtyping to optimize surgical decision-making.

Materials and methods
The retrospective multicenter study was approved by the institutional review boards of the participating institutions, with waiver of informed consent due to anonymized patient data. This study has been reported in line with the STROCSS criteria[15].

Patient selection and data collection
Patients who underwent complete surgical resection for pathologically confirmed stage IA lung adenocarcinoma (LUAD, according to the 8th edition TNM classification) were retrospectively enrolled from three tertiary institutions during distinct time periods: Center 1 (February 2014 to October 2023), Center 2 (September 2016 to November 2022), and Center 3 (September 2015 to October 2023). Clinical data, pathological reports, and preoperative chest CT scans were collected. A development set (training and internal test) was derived from one cohort, while the remaining two cohorts formed external test sets (Fig. 1). Inclusion criteria required: age ≥18 years; histologically confirmed stage IA LUAD; complete surgical excision; preoperative thin-section CT within 8 weeks pre-surgery. Exclusion criteria included incomplete clinical/pathological data or noncompliant CT images (discontinuous/missing/damaged; non-DICOM format).
HIGHLIGHTS
A novel framework using a tabpfn model enables robust ternary classification of stage ia luad subtypes.

The model outperforms conventional machine learning algorithms in a rigorous multicenter validation.

The framework shows potential to complement the accuracy of intraoperative frozen section analysis, especially in challenging cases.

Preoperative subtyping using this model can guide surgical decisions to potentially reduce oncologic undertreatment.

Multicenter validation highlights the model’s generalizability but also reveals challenges, such as performance variability and lower precision for PGL.

Reference standard and histopathological assessment
Histopathological diagnosis served as the reference standard, with all surgical specimens independently reviewed by lung cancer pathologists (K.S., 25 years’ experience). Multiple primary lung cancers (MPLC) diagnoses required multidisciplinary consensus integrating histopathological subtypes and next-generation sequencing to exclude intrapulmonary metastasis[16]. Tumors were classified according to the 2021 WHO criteria as PGL, MIA (≤5 mm stromal invasion), or IAC[2].

Prediction of histogram features using DL-based auxiliary diagnostic system
A fully automated DL-based system (InferRead CT Lung, Infervision, Beijing) was used for nodule segmentation and histogram analysis. This system employs a U-Net architecture trained on over 10 000 manually segmented nodules[17–19], achieving high reproducibility in diameter measurement. Nodule detection, segmentation, and feature extraction were performed without manual correction, as validated in prior clinical studies[20,21]. While proprietary, the system is FDA/CE/NMPA approved and deployed globally, enabling broad applicability of our framework to similar AI platforms.
Based on the detection and segmentation of pulmonary nodules on CT scans, histogram analysis was automatically proceeded. Particularly, quantitative features like long diameter, short diameter, volume, CT values (max, min, median, mean, and standard deviation), solid proportion (>−145 HU), and volume were measured and recorded. Besides, radiomics features including maximum 3D diameter, maximum side area, compactness, energy, kurtosis, skewness, sphericity, entrophy, surface area, and quality of the interested nodules were automatically calculated and recorded for analysis as well (Supplemental Digital Content Figure 1, available at: http://links.lww.com/JS9/G870).
Since one of its algorithms was trained on >1000 chest CT from patients who underwent lesions resection and obtained pathological results, the employed DL system could also predict the malignancy probability scores for the detected pulmonary nodules, which was in strong agreement with histopathology (AUC 0.906, κ = 0.94)[20]. Malignancy probability was also included as another vector of feature. Besides, typical morphological signs could also be recognized and counted, including irregular shape, lobulation, pleural indentation, and spiculation sign. Of note, clinical and demographic features were utilized for further selection and modeling.

Feature selection
Patients from Center 1 were partitioned into training (80%) and internal validation (20%) sets using stratified random sampling. Given the traditional ML models rely on explicit feature selection steps, several feature selection strategies were applied to maximize predictive power and generalizability, including no selection, least absolute shrinkage and selection operator (LASSO) regression selection, and a multi-step selection strategy, in which the univariate analysis (ANOVA/chi-square, P < 0.05) was first used to identify candidate features significantly associated with LUAD histological subtypes, the pairwise Pearson correlations were then calculated to remove redundant features (|r| > 0.9) by retaining the most statistically significant feature (lower P-value) per pair to minimize multicollinearity, followed by the LASSO regression to eventually select features with non-zero coefficients.

Development and evaluation of ternary classification models
In the training phase of our study, several classic machine learning (ML) algorithms were utilized for modeling on training set, including support vector machine (SVM), logistic regression (LR), and multi-layer perceptron (MLP), decision tree (DT) and Random forest (RF). Of note, the pretrained tabular foundation model, TabPFN, was applied in this study for modeling[14].
Ternary classification performance was evaluated by indices of accuracy and macro-averaged area under the receiver operating characteristic (ROC) curve. Meanwhile, its performance on each subtype was evaluated using a range of metrics including sensitivity, specificity, precision, negative prediction value (NPV), F1 score, and AUC, in one-vs-rest approach. Additionally, model performance was benchmarked against intraoperative frozen section analysis across all three centers, with pathological assessments conducted according to institutional protocols.

Statistical analysis
Continuous variables are presented by the means ± standard deviation; categorical variables as frequencies. P < 0.05 was considered statistically significant. In feature selection, significance was examined by ANOVA test for continuous variables and chi-squared tests for category variables. A two-sided 95% confidence interval for AUC was constructed using DeLong test[22]. Cohen’s Kappa coefficient was calculated to measure the agreement between pathology results and model predictions. All statistical analyses were performed with the R statistical package (The R Foundation for Statistical Computing, Vienna, Austria).

Results

Patient cohort characteristics
A total of 583 patients (mean age, 58 years ± 11 [SD]; 386 female) with 770 stage IA LUAD nodules were enrolled across three centers (Center 1: n = 412; Center 2: n = 114; Center 3: n = 57) following application of inclusion and exclusion criteria (Fig. 1). Patients were excluded from the study due to incomplete clinical or pathological records (n = 15), damaged CT images (n = 13), or identification of duplicate cases (n = 6). Demographic and clinical characteristics differed significantly across centers for all variables except age (P < .05), extending to histological subtypes, nodule density, and surgical approaches. Notably, lobectomy rates varied substantially (Center 2: 50.5% [46/91] vs Center 3: 12.5% [7/56]; P < .001). Cohort characteristics are detailed in Table 1.

With a median follow-up of 62 months (IQR: 38–86 months), disease recurrence developed in 17 patients (2.9% of the cohort), and all events occurred exclusively in IAC. Pathological analysis demonstrated that PLC with IAC constituted the predominant recurrence phenotype (70.6%, 12/17), significantly exceeding the recurrence rate of MPLC (29.4%, 5/17); further stratification revealed that all MPLC cases contained invasive adenocarcinoma foci, and 11.8% (2/17) manifested multifocal invasive disease. Center-specific variations were evident: Center 1 accounted for 58.8% (10/17) of recurrences, predominantly involving PLC-IAC (80.0%, 8/10), while Center 2 contributed 41.2% (7/17) with a higher co-occurrence rate of MIA and IAC (42.9%, 3/7). Regarding surgical procedures, 58.8% (10/17) of recurrences followed sublobar resection, specifically segmentectomy (7 cases) or wedge resection (3 cases), underscoring the need for accurate preoperative IAC identification to guide resection extent. For comprehensive recurrence characteristics, see Supplemental Digital Content Table 1, available at: http://links.lww.com/JS9/G871.

Selected feature characteristics
The employed AI system detected all 770 nodules, classifying 676 (87.8%, 676/770) as medium/high malignancy risk, center-level malignancy warning rate was shown in Supplemental Digital Content Figure 2, available at: http://links.lww.com/JS9/G870. Pathologic examination of surgical resection specimens confirmed all nodules as malignant. A total of 32 features were collected from clinical information, histogram features, and radiomics features. Univariate analysis identified 29 features significantly associated with LUAD subtypes (P < 0.05). Pearson correlation analysis for redundancy reduction (PCC > 0.9) resulted in 17 features. LASSO regularization eventually selected 14 features. Meanwhile, direct application of LASSO regression seleted 22 features for modeling. The inter-correlations among the selected features within each set are visualized in the heatmap (Fig. 2 and Supplemental Digital Content Figure 3, available at: http://links.lww.com/JS9/G870).

Three sets of models were trained on all features, LASSO selected features, and multi-step selected features, respectively (Supplemental Digital Content Table 2, available at: http://links.lww.com/JS9/G871). Overall model performance was evaluated and used to select the optimal feature selection strategy. As shown in Supplemental Digital Content Table 3, available at: http://links.lww.com/JS9/G871, multistep selected features enabled an overall better performance of overall accuracy, macro-AUC, and Kappa coefficients for the ternary classification across three testing sets. Detailed further analyses were then performed among models developed based on multi-step selected features.

Performance evaluation and comparison among models on LUAD subtyping
The comparative evaluation of the six models was conducted using macro-AUC and accuracy metrics across three distinct test sets. As shown in Figure 3, all models, with the exception of the DT model, exhibited decent performance on the internal test sets when assessed by either of the aforementioned indices. However, a notable decline in performance was observed on the external test sets, particularly on external test set 2. Among the models compared, TabPFN demonstrated decent performance on exteranl test 1 (accuracy, Macro-AUC, and Kappa Coefficient of 0.789, 0.911, and 0.562) and superior performance on external test sets 2 (optimal accuracy, Macro-AUC, and Kappa Coefficient of 0.672, 0.781, and 0.451, respectively (Supplemental Digital Content Table 3, available at: http://links.lww.com/JS9/G871). ROC curves of each model was presented in Supplemental Digital Content Figure 4, available at: http://links.lww.com/JS9/G870 and statistical comparisons of AUCs for each class were summarized in Supplemental Digital Content Table 4, available at: http://links.lww.com/JS9/G871.

Performance evaluation of TabPFN for ternary classification
Given the optimal overall performance of TabPFN model on ternary classification, an extensive series of evaluations was conducted to further assess its capabilities. In addition to the decent performance on the internal test set as evidenced by the AUC values of 0.972, 0.874, and 0.965 for distinguishing PGL, MIA, and IAC from the others, substantially enhanced robustness was observed in external test sets, especially on external test 2, achieving AUC values of 0.695, 0.750, and 0.897 for PGL, MIA, and IAC (Fig. 4A-C). Besides, the Kappa coefficients of TabPFN model on internal test sets and external test sets 1 and 2 reached 0.7247, 0.5620, and 0.4518 (Fig. 4D-F), indicating good and moderate consistency between predictions and ground truth on internal and external test sets, respectively. Detailed evaluation metrics were summarized in Table 2.

Clinical validation against intraoperative frozen section
Comparative analysis revealed significant center-based disparities in diagnostic accuracy: intraoperative frozen section demonstrated dramatically lower accuracy at Center 3 (43.1%, 25/58) versus Center 1 (87.2%, 41/47) and Center 2 (75.0%, 6/8). By contrast, the TabPFN model employing multi-step selected features demonstrated significantly improved performance (accuracy of 92.3%, 87.5%, and 67.2% in Centers 1–3). Error decomposition confirmed overdiagnosis as the predominant failure pattern (72.5%, 29/40 errors), with misclassification of MIA as IAC accounting for 93.1% (27/29) of overdiagnosis errors (Table 3). Center 3 contributed disproportionately to diagnostic discordance, responsible for 82.8% (24/29) of overdiagnosis cases, of which 95.8% (23/24) were MIA-as-IAC misclassifications. Crucially, borderline lesions between MIA and IAC caused 67.5% (27/40) of all disagreements. This disparity likely stems from selection bias (restricting frozen section to equivocal nodules) and Center 3’s diagnostic inconsistency.

Subgroup analysis of the optimal TabPFN model performance for ternary classification
To delve deeper into the potential influencing factors for the TabPFN model and to identify suitable application cohorts, subgroup analyses were conducted, taking into account factors such as sex, age, number of lesions, and nodule types. The performance across different sex and age groups did not exhibit any significant differences on all test sets (Fig. 5 A B) even though performance on male patients were all relative better than that in female patients on both external test sets. Notably, superior performance was observed in those with a single lesion (Fig. 5C) on external test set 1and in mixed ground-glass nodule (mGGN) compared to pGGN lesions (Fig. 5D) on both external test sets.

Discussion
Precise preoperative differentiation among PGL, MIA, and IAC is imperative for optimizing surgical management of stage IA LUAD, yet remains elusive due to overlapping imaging characteristics. This multicenter investigation establishes that our TabPFN framework, leveraging AI-derived histogram features, achieves robust ternary classification of LUAD subtypes. The model consistently outperformed conventional machine learning algorithms and intraoperative frozen section analysis during rigorous external validation. In specific, TabPFN attained a macro-AUC of 0.781 and an accuracy of 67.2% on external test set 2, exceeding frozen section performance by 24.1% and surpassing the leading traditional model (Random Forest AUC 0.760) by 2.8%. These findings highlight the framework’s capacity to mitigate inter-institutional diagnostic variability.
The TabPFN model’s ability to discriminate among PGL, MIA, and IAC represents a notable improvement beyond existing methodologies. Prior studies predominantly focused on binary classification (e.g., invasive vs. non-invasive)[23], TabPFN demonstrates decent performance (Macro-AUC: 0.781–0.911) in ternary classification even on external test sets, benefiting from its enhanced generalization in limited-sample scenarios achieved via the unique pretraining paradigm which leverages synthetic tabular data to learn universal feature interactions[14]. Of note, a key strength of TabPFN lies in its robust handling of multicollinearity – a persistent challenge in radiomics – through dynamic weighting of discriminative parameters (e.g., nodule compactness and entropy) and suppression of redundant variables (e.g., maximum diameter). Notably, TabPFN demonstrated robust adaptability across diverse feature sets in the determination of optimal feature selection methodology. In contrast, traditional ML models, like LR and MLP, exhibited declined accuracy on external test 2 when using all features for modeling.
Of note, unlike CNN-based models that require GPU acceleration[24], TabPFN operates efficiently on standard CPUs with dynamic weight allocation to emphasize key parameters. In addition, its compatibility with PACS (Picture Archiving and Communication Systems) further enables the seamless integration into radiology workflows – a critical advantage for real-world deployment compared to prior radiomics tools[25]. Of note, the integration of TabPFN with a commercial AI system (InferRead CT Lung) addresses two persistent barriers to clinical adoption: workflow disruption and interoperator variability. By automating feature extraction – including nodule segmentation, histogram analysis, and malignancy probability estimation – the system reduces manual measurement errors that plague traditional radiomics. While manual radiomic assessments historically exhibit moderate reproducibility due to subjective tumor contouring and measurement variability, the standardized algorithms of InferRead CT Lung provide a technical basis for improved consistency in quantifying critical parameters like solid proportion (> −145 HU). This automation not only standardizes inputs but also enables real-time predictions, making the tool viable for both preoperative and intraoperative use. Surgeons could feasibly upload preoperative CT scans, receive instant subclassification results, and adjust resection plans before incision – a workflow unattainable with frozen sections requiring 20–40 minutes per analysis[13].
Clinically, the model’s superiority over intraoperative frozen section analysis (accuracy on external test set 2: 67.2% vs. 43.1% accuracy) carries profound implications. Frozen sections, which were considered as the gold standard for intraoperative decision-making, suffered from sampling errors and interobserver variability, particularly in heterogeneous lesions[26]. By analyzing entire nodule volumes quantitatively, TabPFN helps overcome these limitations, offering surgeons a reproducible tool to guide resection margins. Our recurrence data showed that 38.5% (5/13) of IAC recurrences patients underwent sublobar resection without intra-operative frozen sections analyses. High IAC detection capability (AUC: 0.897-0.965) of TabPFN made the redirection of sublobar resection candidates to lobectomy when appropriate possible. For example, among 111 IAC lesions which did not undergo frozen section analyses at Center 2, the model correctly identified 40 out of 45 leisons which underwent sublobar resections as IAC when serving as the intraoperative subtyping tool at expenses of two lobectomy-resected lesions being recognized as MIA (2/66). This aligns with JCOG0802 principles supporting parenchymal-sparing resection only when invasiveness is definitively excluded[27]. Future prospective trials should evaluate how TabPFN-guided decisions impact surgical outcomes.
The clinical relevance of these findings is amplified by sex- and nodule-type-specific performance trends. Relative higher accuracy in male patients (85.6% vs. 73.7%) may reflect biological differences in tumor microenvironment composition, such as higher fibroblast activation protein (FAP) expression in males, which enhances stromal invasion visibility on CT[28]. Similarly, the model’s significantly higher accuracy in mixed ground-glass nodules (averaged accuracy 0.842 vs 0.590 on all test sets) aligns with histopathological evidence that solid components correlate with invasiveness[29], whereas lower performance in pGGNs underscores their radiomic homogeneity. Subgroup analyses provided insights for potential applicable cohorts: clinicians might prioritize TabPFN for mGGNs in males while combining it with liquid biopsy or interval CT surveillance for pure GGNs. While further validation on more enlarged external test sets were warranted to verify these findings.
Despite its strengths, this study has limitations that warrant consideration. First, the retrospective design introduces selection bias, as only resected nodules were included. Indolent lesions managed with surveillance – which may exhibit distinct radiomic profiles – were excluded, potentially inflating model performance. Prospective trials incorporating non-surgical cases are needed to validate real-world utility. Second, the model’s lower precision for PGL classification suggests precursor lesions lack distinct radiomic signatures, necessitating histopathological correlation for definitive diagnosis. Future iterations could integrate cytological features from biopsy specimens to enhance PGL detection. Finally, the absence of molecular data (e.g., EGFR, KRAS mutations) limits our ability to correlate radiomic features with tumor biology – a critical area for investigation given emerging links between EGFR mutations and ground-glass opacity patterns[9].

Conclusions
This multicenter study developed a promising tool for preoperative LUAD subtyping using TabPFN algorithm. Based on automated AI-driven feature extraction and the robust performance of TabPFN, our framework helps address the reproducibility issue in medical AI while providing a blueprint for real-world implementation. The findings highlight the potential of foundation models in radiomic analysis, particularly when coupled with standardized commercial systems. As thoracic oncology embraces precision strategies, tools based on TabPFN shares great potentials to aligning diagnostic accuracy with therapeutic innovation, ultimately improving outcomes for patients with early-stage LUAD.