Survival risk prediction for patients with lung adenocarcinoma containing micropapillary components based on weakly supervised pathomics

Abstract

Background

The presence of a micropapillary component is an independent adverse prognostic predictor for patients with lung adenocarcinoma. This study aimed to develop a pathomics model based on weakly supervised deep learning to accurately predict postoperative survival risk in patients with lung adenocarcinoma containing micropapillary components by analyzing whole slide images.

Methods

Hematoxylin and eosin (H&E) stained whole slide images from 202 patients with lung adenocarcinoma containing micropapillary components who underwent surgical resection were retrospectively collected. A weakly supervised learning framework was adopted, utilizing ResNet18 and vision transformer (ViT) architectures to extract morphological features at the tile level, which were then aggregated to the patient level via multi-instance learning. Least absolute shrinkage and selection operator (LASSO) regression was employed to screen key pathomic features and construct a pathomics prediction model, which was compared with traditional models based on clinicopathological factors. Ultimately, the optimal pathomic features and clinically independent risk factors were integrated to establish a combined predictive nomogram.

Results

Among the deep learning models, ResNet18 demonstrated superior feature extraction capability, with area under the curve (AUC) values of 0.934 and 0.653 in the training and test sets, respectively, outperforming the ViT model. Consequently, the ResNet18 model was selected, yielding 206 extracted features. Following LASSO regression, four key pathomic features were retained. A logistic regression (LR) model built upon these features achieved an AUC of 0.892 in the training set. Univariate and multivariate analyses of clinical variables identified eight significant features, including micropapillary and solid components, for modeling. The clinical feature model constructed using LR achieved AUCs of 0.872 and 0.698 in the training and test sets, respectively. The final combined clinicopathological nomogram significantly improved predictive performance in the training set (AUC =0.945), surpassing both the standalone pathomics and clinical models, and maintained reasonable discriminative ability in the test set (AUC =0.687). Visual analysis using Gradient-Weighted Class Activation Mapping (Grad-CAM) indicated that the model’s decisions were focused on regions rich in micropapillary structures, enhancing interpretability.

Conclusions

Weakly supervised deep learning methods can effectively identify digital pathological features associated with prognosis. A combined model integrating pathomic features with clinical information demonstrated stronger predictive power than models based solely on pathology or clinical data, offering a potential auxiliary tool for individualized postoperative risk stratification in patients with lung adenocarcinoma containing micropapillary components.

Highlight box.

Key findings

• This study developed and validated a weakly supervised pathomics model to predict postoperative survival risk in patients with lung adenocarcinoma containing micropapillary components.

• A combined nomogram integrating deep learning-derived pathomic features with key clinical risk factors [e.g., micropapillary/solid/acinar component proportions, preoperative carcinoembryonic antigen (CEA)] demonstrated superior predictive performance [area under the curve (AUC) =0.945 in training] compared to models using either data type alone.

• ResNet18 outperformed ViT for feature extraction, with model attention focused on micropapillary-rich regions, enhancing interpretability.

What is known and what is new?

• The micropapillary component is a recognized adverse prognostic factor in lung adenocarcinoma. Supervised deep learning in computational pathology often requires exhaustive annotations, limiting clinical scalability.

• We present a specialized, weakly supervised pathomics pipeline requiring only slide-level labels. By fusing automatically extracted digital pathology features with clinical variables, we provide a more accurate and practical tool for risk stratification in this high-risk subtype, moving beyond pan-cancer or broad non‑small cell lung cancer (NSCLC) models.

What is the implication, and what should change now?

• This approach offers an objective, scalable auxiliary tool for identifying high-risk patients early, potentially guiding personalized postoperative surveillance and adjuvant therapy decisions.

• The model requires external validation in multicenter cohorts. Future work should integrate molecular data (e.g., EGFR/ALK) to build a comprehensive multi-omics prognostic system, further bridging digital pathology with clinical decision-making.

Introduction

Lung cancer represents one of the most prevalent and lethal malignancies worldwide, imposing a significant burden on global health (1). According to the fifth edition of the World Health Organization (WHO) Classification of Thoracic Tumours (2021), invasive non-mucinous lung adenocarcinoma is composed of five major histologic growth patterns (2). These patterns should be quantified in pathology reports using percentage increments (typically 5% or 10%). All histological components should be incorporated into the study. However, Lung adenocarcinoma containing a micropapillary component, recognized as a highly invasive subtype of non-small cell lung cancer (NSCLC), exhibits characteristic histopathological features. These include papillary tumor cell clusters devoid of fibrovascular cores, which may detach from or remain attached to alveolar walls, often appearing as ring-like glandular structures floating within alveolar spaces (3,4). Tumor cells are typically small and cuboidal, frequently demonstrating vascular or stromal invasion, with occasional psammoma bodies. The presence of this subtype is strongly associated with intrapulmonary spread, pleural invasion, lymph-node metastasis, and poor patient prognosis (5,6).

Currently, the clinical management of lung adenocarcinoma with a micropapillary component faces several major challenges. First, conventional imaging modalities such as computed tomography (CT) or positron emission tomography-computed tomography (PET-CT) struggle to distinguish micropapillary morphology from true papillary patterns. Second, the high heterogeneity of the tumor microenvironment, coupled with the lack of reliable biomarkers for the micropapillary component, hinders accurate prognostic assessment and treatment stratification. Consequently, developing an objective and accurate tool to predict postoperative survival risk in this patient population holds considerable clinical significance.

Presently, quantification of the micropapillary component relies on pathological evaluation. Hematoxylin and eosin (H&E) staining remains the gold-standard diagnostic method due to its cost-effectiveness and high reproducibility (7-9). In recent years, the integration of digital pathology and artificial intelligence (AI) has created new opportunities for cancer prognosis assessment. Whole-slide imaging (WSI) enables high-throughput digitization of pathological slides (10,11), while deep-learning models can automatically extract prognostically relevant histopathological features from large-scale image datasets, forming a “pathomics” analytical framework (12,13). Prior studies have applied deep learning to survival prediction in various malignancies, including lung cancer (14), gastric cancer (15), glioma (16), and colorectal cancer lung metastases (17).

However, most existing studies focus on pan-cancer prognostic modeling or broad NSCLC cohorts, leaving a gap in specialized predictive tools for the high-risk subtype of lung adenocarcinoma with a micropapillary component. Moreover, current methods predominantly rely on supervised learning, requiring extensive pixel-level annotations that limit clinical applicability (18). Weakly supervised deep-learning approaches (19), which require only slide-level labels for feature learning and model development, are better suited for analyzing large-scale clinical histopathology datasets. Therefore, this study aims to construct a pathomics model based on weakly supervised deep learning. By analyzing H&E-stained whole-slide images, the model will automatically extract relevant pathological features and integrate them with clinical variables to establish a combined predictive nomogram. This approach seeks to improve the accuracy of postoperative survival risk prediction for patients with lung adenocarcinoma containing a micropapillary component, thereby providing a supportive tool for individualized therapeutic decision-making. We present this article in accordance with the TRIPOD reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1-2647/rc).

Methods

Patient cohort and selection criteria

This retrospective study included 213 patients diagnosed with lung adenocarcinoma containing a micropapillary component who underwent surgical resection at Jiaxing First Hospital between January 2018 and October 2023. The follow-up period ended in April 2025. The study was approved by the Ethics Review Committee of the Affiliated Hospital of Jiaxing University (Jiaxing First Hospital) (approval No. 2025-KY-411). This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. All patients have signed the informed consent form. Inclusion criteria were as follows: (I) pathologically confirmed primary lung adenocarcinoma with micropapillary structures; (II) availability of digitized WSIs of H&E-stained pathological sections; (III) no history of neoadjuvant therapy or radiotherapy prior to surgery; (IV) complete clinicopathological and follow-up data. Exclusion criteria included: (I) history of other malignancies or development of secondary cancers postoperatively; (II) metastatic lung cancer; (III) loss to follow-up; (IV) poor-quality or blurred pathological images. After applying these criteria, 11 patients were excluded, resulting in a final cohort of 202 patients. The cohort was randomly divided into a training set (n=141) and a test set (n=51) in a 7:3 ratio.

Clinical data collection

Clinical and pathological data were systematically collected, including patient demographics (age, sex), surgical procedure, tumor location, tumor size, histological grade, and postoperative pathological findings such as the presence of vascular tumor thrombus, pleural invasion, perineural invasion, lymph node metastasis, and the proportion of micropapillary component. Pathological assessment was performed according to the 2021 WHO classification criteria. Tumor staging followed the 8th edition Union for International Cancer Control/American Joint Committee on Cancer (UICC/AJCC) tumor, node, metastasis (TNM) staging system. Additional data included genetic testing results, postoperative adjuvant therapy, and follow-up information on survival, recurrence, and metastasis. Smoking status was defined as a documented history of tobacco use exceeding one year, irrespective of current or former usage. Overall survival (OS) was defined as the interval from surgery to death or last follow-up. Follow-up protocol included chest CT scans and tumor marker assessments at 3 months, 6 months, and 1 year postoperatively, with subsequent evaluations every 6 to 12 months based on clinical status. Telephone follow-up was conducted to document symptoms, imaging findings, and laboratory results.

Slide preparation and image acquisition

Surgical specimens were fixed in 10% formalin for 12 hours, embedded in paraffin, and sectioned at 4 µm thickness. Slides were stained with H&E for pathological evaluation. A JF Bio KF-PRO-400 slide scanner was used to digitize all H&E slides at 16× magnification, saved in KFB format. Images were subsequently converted to SVS format using a KFB-to-SVS conversion plugin. Each digital slide was independently reviewed by two experienced pathologists (with 3 and 8 years of expertise in lung cancer pathology, respectively). Discrepancies were resolved through consultation with a third senior pathologist.

Pathomics workflow

The analytical pipeline comprised four main stages: slide preprocessing, model training, feature extraction, and model construction.

Slide preprocessing

WSIs were segmented into non-overlapping 512×512 pixel patches using the OKT-WSI2patch tool. Background patches were filtered out using the OnekeyAI CNN patch2predict toolkit to enhance computational efficiency. Color normalization was performed using the Reinhard method to minimize staining variation. Subsequently, Z-score normalization was applied to each RGB channel to standardize intensity distributions. The processed patches served as input for deep learning models.

Model training

Two deep learning architectures were employed: ResNet18 (20) and ViT (21). Training was optimized using stochastic gradient descent with a cosine annealing learning rate schedule, defined as:

$${\eta }_t={\eta }_{min}^i+\frac{1}{2}\left({\eta }_{max}^i-{\eta }_{min}^i\right)\left(1+\cos \left(\frac{T_{cur}}{T_i}\pi \right)\right)$$ [1]

In this formula, ${\eta }_{min}^i=0$ sets the minimum learning rate, ${\eta }_{max}^i=0:01$ establishes the maximum learning rate, and $T_i=50$ represents the number of iterations during the iterative training process. This learning rate plan adopts a gradual reduction strategy and achieves precise model refinement throughout the training phase. To further improve the training method and enhance the prediction accuracy, we employ stochastic gradient descent as an optimization technique. In addition, the softmax cross-entropy, as a loss function, is helpful for calculating the probability distribution of the expected target category

Feature extraction and model construction

Following model training, patch-level predictions and corresponding probabilities were generated. Two aggregation methods were applied to derive slide-level representations: (I) patch likelihood histogram (PLH), which quantifies the distribution of prediction probabilities across patches within a WSI; and (II) bag-of-words (BoW) with term frequency-inverse document frequency (TF-IDF) weighting, generating a feature vector per WSI. These features were then used to train traditional machine learning classifiers.

Feature selection was performed using LASSO regression to identify non-redundant, prognostic pathomic features. The selected features were used to build a pathomics prediction model. We use machine learning models such as LR, support vector machine (SVM), K-nearest neighbors (KNN), random forest, ExtraTrees, extreme gradient boosting (XGBoost), lightweight gradient booster (LightGBM), and multi-layer perceptron (MLP) for modeling. Concurrently, univariate and multivariate analyses were conducted to identify clinical risk factors associated with survival risk. For the multivariate LR model, variables with P<0.05 in the univariate analysis were initially considered. A forward stepwise selection procedure was then applied to identify the final independent predictors. A combined nomogram integrating significant pathomic and clinical predictors was developed using LR.

Model evaluation

Model performance was assessed using the area under the curve (AUC). Calibration was evaluated via calibration curves and the Hosmer-Lemeshow test. Clinical utility was quantified using decision curve analysis (DCA). DeLong’s test was used to compare AUCs between models.

Statistical analysis

Continuous variables were compared using Student’s t-test or Mann-Whitney U test, as appropriate. Categorical variables were analyzed using the chi-square test. All data analyses were performed using Python 3.7.12 on the OnekeyAI platform version 4.9.1. Statistical evaluation was conducted using Statsmodels version 0.13.2. The implementation of machine learning algorithms (including SVMs) was accomplished using Scikit-learn version 1.0.2. Our deep learning framework is developed based on PyTorch version 1.11.0 and optimized through CUDA version 11.3.1 and cuDNN version 8.2.1 to enhance performance. A two-sided P value <0.05 was considered statistically significant.

Results

Baseline characteristics and clinical feature selection

A total of 202 patients with lung adenocarcinoma containing a micropapillary component were included in the final analysis. The mean age was 64.26±9.02 years, and 60.4% of patients were female. The majority (71.78%) presented with pathological stage I disease. The mean tumor diameter was 2.29±0.99 cm, and the average proportion of the micropapillary component was 11.52%±13.30%. One stage IV patient was included in the study cohort. Although preoperative imaging showed no evidence of metastasis, pleural dissemination was identified during surgery and confirmed as pleural metastasis by postoperative pathology. Since the patient met the predefined inclusion criteria at the time of enrollment, they were retained in the analysis. All baseline characteristics are summarized in Table 1. The training and test sets were well-balanced across most features, confirming the comparability of the dataset split.

Table 1. Baseline characteristics of all cohorts.

Characteristic	Overall (n=202)	Training (n=141)	Test (n=61)	P
Age (years)	64.26±9.02	64.06±8.85	64.72±9.46	0.62
Tumor size (cm)	2.29±0.99	2.18±0.96	2.53±1.02	0.02
Lepidic (%)	1.010±5.840	1.100±6.240	0.820±4.850	0.75
Acinar (%)	45.160±29.310	45.650±29.120	44.050±29.980	0.65
Papillary (%)	30.270±25.780	31.000±25.560	28.570±26.420	0.43
Solid (%)	6.680±16.620	6.280±15.740	7.610±18.590	0.54
Micropapillary (%)	11.520±13.300	10.760±11.850	13.300±16.140	0.61
Complex glandular (%)	6.340±16.530	6.170±15.820	6.720±18.210	0.69
Gender				0.84
Female	122 (60.400)	84 (59.574)	38 (62.295)
Male	80 (39.600)	57 (40.426)	23 (37.705)
CEA				0.70
Negative	151 (74.752)	107 (75.887)	44 (72.131)
Positive	51 (25.248)	34 (24.113)	17 (27.869)
Smoking				0.01
No	171 (84.653)	113 (80.142)	58 (95.082)
Yes	31 (15.347)	28 (19.858)	3 (4.918)
Chemotherapy				0.98
No	174 (86.139)	122 (86.525)	52 (85.246)
Yes	28 (13.861)	19 (13.475)	9 (14.754)
Surgical method				0.051
Lobectomy	135 (66.832)	91 (64.539)	44 (72.131)
Segmentectomy	34 (16.832)	28 (19.858)	6 (9.836)
Wedge	27 (13.366)	20 (14.184)	7 (11.475)
Lobectomy + wedge	5 (2.475)	1 (0.709)	4 (6.557)
Segmentectomy + wedge	1 (0.495)	1 (0.709)	0
Lymph node positive				0.15
No	151 (74.752)	110 (78.014)	41 (67.213)
Yes	51 (25.248)	31 (21.986)	20 (32.787)
TNM stage				0.45
Stage I	145 (71.782)	105 (74.468)	40 (65.574)
Stage II	24 (11.881)	14 (9.929)	10 (16.393)
Stage III	32 (15.842)	21 (14.894)	11 (18.033)
Stage IV	1 (0.495)	1 (0.709)	0
T stage				0.03
T1	155 (76.733)	112 (79.433)	43 (70.492)
T2	42 (20.792)	27 (19.149)	15 (24.590)
T3	3 (1.485)	0	3 (4.918)
T4	2 (0.990)	2 (1.418)	0
N stage				0.11
N0	151 (74.752)	110 (78.014)	41 (67.213)
N1	20 (9.901)	10 (7.092)	10 (16.393)
N2	31 (15.347)	21 (14.894)	10 (16.393)
M stage				>0.99
M0	201 (99.505)	140 (99.291)	61 (100.000)
M1	1 (0.495)	1 (0.709)	0 (0.000)
STAS				>0.99
No	151 (74.752)	105 (74.468)	46 (75.410)
Yes	51 (25.248)	36 (25.532)	15 (24.590)
Perineural invasion				>0.99
No	197 (97.525)	137 (97.163)	60 (98.361)
Yes	5 (2.475)	4 (2.837)	1 (1.639)
Vascular tumor thrombus				0.66
No	160 (79.208)	110 (78.014)	50 (81.967)
Yes	42 (20.792)	31 (21.986)	11 (18.033)

Data are presented as n (%) or mean ± standard deviation. CEA, carcinoembryonic antigen; STAS, spread through air spaces; TNM, tumor, node, metastasis.

The distribution of growth pattern components within the cohort was as follows {median [interquartile range (IQR)]}: micropapillary 5% (2–15%), solid 5% (0–10%), acinar 45% (20–70%), papillary 10% (0–30%). Notably, lepidic components were absent (0%) in 83.7% of cases (169/202), which aligns with the recognized aggressive, poorly differentiated nature of lung adenocarcinoma with micropapillary components.

Table S1 showed Spearman correlation analysis that the proportion of micropapillary component was positively correlated with the solid component proportion (r=0.238, P=0.001). In contrast, it was inversely correlated with both acinar (r=−0.162, P=0.02) and papillary components (r=−0.144, P=0.04). No significant correlation was observed with lepidic components.

The micropapillary proportion was significantly higher in patients with spread through air spaces (STAS) compared to those without STAS (mean ± SD: 16.3%±18.0% vs. 9.7%±11.2%; Mann-Whitney U test, P=0.003). Similarly, it was elevated in patients with lymph-node metastasis relative to node-negative cases (15.6%±17.3% vs. 9.8%±11.6%; P=0.008). No significant association was detected between micropapillary proportion and patient gender or smoking history. All the details were revealed in the Table S2.

Univariate and multivariate analyses were performed to identify clinicopathological factors associated with postoperative survival (Table 2). Univariate analysis identified age, proportion of solid component, proportion of acinar component, presence of complex glandular structures, proportion of papillary component, proportion of micropapillary component, tumor size, postoperative chemotherapy, TNM stage, preoperative carcinoembryonic antigen (CEA) level, surgical method, T stage, vascular tumor thrombus, smoking history, gender, and STAS as significant risk factors (P<0.05). Subsequent multivariate analysis confirmed the following independent risk factors: age, proportion of solid component, proportion of acinar component, complex glandular structures, proportion of papillary component, proportion of micropapillary component, postoperative chemotherapy, and preoperative CEA level.

Table 2. Univariate and multivariate analysis of clinical features.

Variable	Univariate analysis				Multivariate analysis
	OR	95% CI	P value		OR	95% CI	P value
Lepidic	0.987	0.943–Inf	0.64		–	–	–
Age	1.033	1.026–Inf	<0.001		0.900	0.856–0.946	0.001
Solid	1.035	1.009–Inf	0.02		1.076	1.017–1.138	0.03
Acinar	1.042	1.030–Inf	<0.001		1.126	1.067–1.188	<0.001
Complex glandular	1.065	1.024–Inf	0.008		1.114	1.050–1.181	0.002
Papillary	1.087	1.061–Inf	<0.001		1.130	1.099–1.162	<0.001
Micropapillary	1.130	1.084–Inf	<0.001		1.132	1.084–1.182	<0.001
N stage	1.748	1.185–Inf	0.02		–	–	–
Tumor size	2.252	1.850–Inf	<0.001		0.594	0.282–1.250	0.25
Chemotherapy	2.800	1.189–Inf	0.048		0.124	0.026–0.582	0.03
Perineural invasion	3.000	0.449–Inf	0.34		–	–	–
TNM stage	3.244	2.387–Inf	<0.001		5.204	0.614–44.080	0.20
CEA	3.250	1.672–Inf	0.004		0.064	0.031–0.133	<0.001
Lymph node positive	3.428	1.692–Inf	0.004		1.380	–	–
Surgical method	4.258	2.995–Inf	<0.001		0.992	0.498–1.974	>0.99
T stage	4.484	3.149–Inf	<0.001		0.364	0.062–2.143	0.35
Pleural invasion	6.200	2.807–Inf	<0.001		1.080	0.209–5.596	0.94
Vascular tumor thrombus	6.750	2.795–Inf	<0.001		1.067	0.219–5.191	0.95
Smoking	8.334	3.050–Inf	0.001		0.357	0.029–4.349	0.50
Gender	10.400	4.816–Inf	<0.001		6.059	0.751–48.911	0.16
STAS	17.000	5.135–Inf	<0.001		1.274	0.198–8.207	0.83
M stage	2.619	0.000–Inf	>0.99		–	–	–

CEA, carcinoembryonic antigen; CI, confidence interval; OR, odds ratio; STAS, spread through air spaces; TNM, tumor, node, metastasis.

Using these eight independent risk factors, multiple machine learning models were trained (Table S3 for details). The LR model demonstrated optimal balanced performance, achieving an AUC of 0.872 [95% confidence interval (CI): 0.790–0.953] in the training set. In the test set, the model maintained an AUC of 0.698 (95% CI: 0.525–0.871), indicating reasonable generalizability.

The median OS was 39 months (IQR, 24–64 months), with 1-, 3-, and 5-year OS rates of 94.2%, 72.5%, and 58.3%, respectively. When stratified by micropapillary component proportion, median OS decreased progressively: 64 months for Group 1 (IQR, 0–5%), 39 months for Group 2 (IQR, 6–20%), 28 months for Group 3 (IQR, 21–50%), and 19 months for Group 4 (IQR, 51–100%) (Figure 1).

Figure 1.

Kaplan-Meier survival analysis stratified by micropapillary component proportion. Group 1: 0–5% (n=85, median OS 64 months), Group 2: 6–20% (n=65, median OS 39 months), Group 3: 21–50% (n=35, median OS 28 months), Group 4: 51–100% (n=17, median OS 19 months). Log‑rank test P<0.05. OS, overall survival.

Deep learning model training and feature extraction

Two deep learning architectures, ResNet18 and ViT, were trained for feature extraction. ResNet18 demonstrated excellent discriminative ability in the training set, with an AUC of 0.934 (95% CI: 0.934–0.935), significantly outperforming the ViT model (AUC =0.752) (Figure 2A). In the test set, ResNet18 maintained a generalization AUC of 0.653, while the ViT model achieved an AUC of 0.604 (Figure 2B), underscoring the advantage of convolutional neural networks (CNNs) in capturing local histomorphological patterns relevant to micropapillary structures.

Figure 2.

Performance of deep learning models and selection of pathomic features. (A) ROC curve comparison of the ResNet18 model in the training sets. (B) ROC curve comparison of the ViT model in the test sets. (C,D) LASSO coefficient paths and cross‑validation error curve. (E) The four key pathomic features selected by LASSO regression and their corresponding coefficients. AUC, area under the curve; CI, confidence interval; LASSO, least absolute shrinkage and selection operator; ROC, receiver operating characteristic; ViT, vision transformer.

From the whole-slide images, the two deep learning architectures extracted a total of 206 pathomic features. Dimensionality reduction via LASSO regression yielded four non-zero coefficient features, which were retained for subsequent modeling (Figure 2C-2E).

Pathomics model development and evaluation

The performance of multiple machine learning classifiers built upon the selected pathomic features was compared (Table 3). The LR model again showed the most balanced performance, with an AUC of 0.892 in the training set. However, its performance in the test set was modest (AUC =0.566), suggesting limited independent generalizability of the pathomics features alone.

Table 3. Machine learning modeling of pathological features.

Model	Cohort	Accuracy	AUC (95% CI)	Sensitivity	Specificity	PPV	NPV
LR	Train	0.695	0.892 (0.824–0.960)	0.661	1.000	1.000	0.246
	Test	0.883	0.566 (0.287–0.845)	0.962	0.286	0.911	0.500
SVM	Train	0.879	0.673 (0.471–0.874)	0.913	0.571	0.951	0.421
	Test	0.867	0.458 (0.175–0.741)	0.943	0.286	0.909	0.400
KNN	Train	0.099	0.896 (0.845–0.947)	0.000	1.000	0.000	0.099
	Test	0.117	0.594 (0.384–0.805)	0.000	1.000	0.000	0.117
Random forest	Train	0.972	0.999 (0.998–1.000)	0.969	1.000	1.000	0.778
	Test	0.750	0.518 (0.327–0.708)	0.830	0.143	0.880	0.100
ExtraTrees	Train	0.099	1.000 (1.000–1.000)	0.000	1.000	0.000	0.099
	Test	0.700	0.600 (0.377–0.823)	0.736	0.429	0.907	0.176
XGBoost	Train	0.965	0.994 (0.986–1.000)	0.961	1.000	1.000	0.737
	Test	0.117	0.566 (0.368–0.764)	0.000	1.000	0.000	0.117
LightGBM	Train	0.823	0.927 (0.873–0.982)	0.819	0.857	0.981	0.343
	Test	0.700	0.621 (0.382–0.860)	0.736	0.429	0.907	0.176
MLP	Train	0.660	0.877 (0.802–0.952)	0.622	1.000	1.000	0.226
	Test	0.900	0.590 (0.305–0.875)	0.981	0.286	0.912	0.667

AUC, area under the curve; CI, confidence interval; KNN, K-nearest neighbors; LightGBM, lightweight gradient booster; LR, logistic regression; MLP, multi-layer perceptron; NPV, negative predictive value; PPV, positive predictive value; SVM, support vector machine; XGBoost, extreme gradient boosting.

To enhance model interpretability, Gradient-Weighted Class Activation Mapping (Grad-CAM) was applied to visualize the regions of focus within WSIs for the ResNet18 model (Figure 3). The heatmaps demonstrated that the model spontaneously attended to areas of high tumor cellularity and regions rich in micropapillary structures during prediction, aligning well with established pathological landmarks.

Figure 3.

Grad-CAM visualization for the ResNet18 model. (A) Preprocessed patches (H&E staining, ×16). (B) The red regions highlight the areas of focus identified by the model. (C) Validation of whether the model accurately identifies and localizes the target. Grad-CAM, Gradient-Weighted Class Activation Mapping; H&E, hematoxylin and eosin.

Development and evaluation of the combined model

To improve predictive performance, the selected pathomic features were integrated with the clinical independent risk factors to construct a multimodal combined nomogram (Figure 4). The nomogram incorporates common risk predictors such as solid component proportion, micropapillary component proportion, and complex glandular structures, assigning a score to each variable. The total score corresponds to an individual patient’s estimated postoperative survival risk. To use the nomogram, first locate the patient-specific values on each predictor axis. Draw vertical lines upward to the “Points” axis to determine the contribution of each variable. Sum all point values to obtain the total score, then project the total score downward to the bottom risk-probability axis to read the corresponding probability of high postoperative risk. For example, a 55-year-old patient with 20% micropapillary component, 10% solid component, elevated CEA, present complex glandular structures, and receiving chemotherapy would receive the following point contributions: age 12 points, micropapillary component 60 points, solid component 35 points, elevated CEA 130 points, complex glandular structures 50 points, and chemotherapy—20 points. The total score of 267 points corresponds to a high-risk probability of 78%.

Figure 4.

Nomogram of the combined prediction model. CEA: 0, negative; 1, positive; chemotherapy: 0, no; 1, yes. CEA, carcinoembryonic antigen; path, path scores.

The combined model demonstrated the highest predictive efficacy in the training set (Table 4), with an AUC of 0.945 (Figure 5A), outperforming both the standalone clinical model (AUC =0.872) and the pathomics model (AUC =0.892). In the test set, the combined model achieved an AUC of 0.687 (Figure 5B), which was marginally lower than the clinical model’s AUC of 0.698. While DeLong’s test showed significant improvement over the clinical model in training (P=0.04), this difference was not statistically significant in testing (P=0.052) (Figure 5C,5D). DCA showed that the combined model provided superior net clinical benefit across a wide range of threshold probabilities compared to “treat-all” or “treat-none” strategies, supporting its potential clinical utility (Figure 5E,5F). This confirms that the integrated pathomic features provide complementary prognostic information, enhancing the overall predictive capability.

Table 4. Model comparisons.

Cohort	Model	Accuracy	AUC (95% CI)	Sensitivity	Specificity	PPV	NPV
Train	Clinical	0.730	0.872 (0.790–0.953)	0.709	0.929	0.989	0.260
	Pathomics	0.695	0.892 (0.824–0.960)	0.661	1.000	1.000	0.246
	Combined	0.844	0.945 (0.905–0.985)	0.827	1.000	1.000	0.389
Test	Clinical	0.550	0.698 (0.525–0.871)	0.491	1.000	1.000	0.206
	Pathomics	0.883	0.566 (0.287–0.845)	0.962	0.286	0.911	0.500
	Combined	0.467	0.687 (0.494–0.881)	0.396	1.000	1.000	0.179

AUC, area under the curve; CI, confidence interval; NPV, negative predictive value; PPV, positive predictive value.

Figure 5.

Evaluation and comparison of the pathomics model, clinical model, and combined model. (A) ROC curves of the three models in the training set. (B) ROC curves of the three models in the test set. (C) DeLong’s test in the training set. (D) DeLong’s test in the test set. (E) DCA in the training set. (F) DCA in the test set. AUC, area under the curve; CI, confidence interval; DCA, decision curve analysis; ROC, receiver operating characteristic.

Discussion

Lung cancer continues to be the leading cause of global cancer-related morbidity and mortality, with over 2 million new cases and approximately 1.8 million deaths annually (22). Despite considerable advances in diagnostic and therapeutic strategies (23), the overall survival of patients with lung adenocarcinoma remains suboptimal, largely attributable to the disease’s pathological heterogeneity and diverse clinical behavior. The presence of a micropapillary component is a well-recognized adverse prognostic factor, with studies consistently demonstrating that patients with this subtype exhibit significantly worse overall and disease-free survival compared to other histological patterns (4).

Within the context of lung adenocarcinomas containing micropapillary components, our analysis indicates that the proportion of micropapillary pattern maintains a significant association with poor outcome even after adjustment for other clinicopathological variables. However, without a control group of adenocarcinomas lacking micropapillary components, we cannot definitively establish whether this represents a truly independent prognostic effect specific to micropapillary morphology or reflects broader associations common to high-grade adenocarcinoma patterns. Survival analysis indicated that patients with a higher proportion of micropapillary component had a markedly reduced 5-year survival rate. Baseline characteristics also revealed that a greater micropapillary content was frequently associated with other adverse pathological features, including STAS and complex glandular structures, collectively contributing to poor postoperative outcomes. Notably, the majority of patients (71.78%) in our cohort presented with stage I disease, highlighting the critical need for precise identification of high-risk pathological patterns—even in early-stage cases—to inform appropriate management strategies. By integrating digitally extracted pathomic features with established clinical risk factors, our model demonstrated an enhanced ability to stratify high-risk patients, potentially facilitating more personalized postoperative surveillance and adjuvant therapy decisions.

Two primary deep learning paradigms were considered for this study: supervised and weakly supervised learning. Fully supervised learning requires exhaustive, pixel-level annotations, wherein pathologists manually delineate each region of interest. This approach provides the AI with a detailed instruction manual, enabling accurate identification of specific morphological features. However, it is costly, time-consuming, and subject to inter-observer variability, thereby limiting scalability for large-scale retrospective studies. Its strengths lie in precise localization and strong interpretability, making it suitable for well-defined tasks such as mitotic counting or the detection of particular structures. Weakly supervised learning relies solely on slide-level labels without any pixel-wise annotations. Under a multiple-instance learning framework, the whole-slide image is treated as a “bag” containing hundreds to thousands of image patches, and the model learns to identify which regions are most predictive of the clinical outcome. This paradigm is akin to showing the AI patient outcomes and allowing it to discover prognostic patterns autonomously. It aligns better with clinical workflows, offers superior scalability, and can uncover novel prognostic features beyond predefined morphological criteria. It is particularly suited for studies involving histologically heterogeneous and prognostically complex entities, such as the prognostic modeling of lung adenocarcinoma with micropapillary components in the present investigation (24). Prior studies have successfully applied both paradigms. For instance, Coudray et al. employed supervised deep learning to classify lung cancer subtypes and predict EGFR mutations from frozen or paraffin-embedded tissue sections, achieving an AUC of 0.97 (18). However, their work did not specifically focus on micropapillary components or incorporate H&E-stained slides into prognostic modeling. Similarly, Han et al. utilized a weakly supervised approach with DenseNet121, ResNet50, and Inception V3 to predict major pathological response following neoadjuvant chemoimmunotherapy in NSCLC (AUC =0.998) (25). While their study integrated pathological and clinical features, it did not evaluate long-term survival outcomes. Our work extends these efforts by establishing a direct link between deep learning-derived pathomic signatures and postoperative survival risk in a high-risk adenocarcinoma subtype, thereby strengthening the clinical relevance of such models.

The substantial performance attenuation of the pathomics-alone model (training AUC 0.892 → test AUC 0.566) indicates that the AI-derived features in their current form may not generalize robustly beyond the training cohort. This highlights a key challenge in applying weakly-supervised deep learning to rare histological subtypes where sample sizes are limited.

Regarding model architecture selection, our results indicate that ResNet18 outperformed ViT in this specific task. This may be attributed to the inherent strength of CNNs in capturing localized histopathological features—such as cellular arrangement and stromal interaction—that are critical for identifying micropapillary structures. Nevertheless, the potential of Transformer-based models in modeling long-range dependencies within tissue architecture should not be overlooked. Future studies could explore hybrid CNN-Transformer architectures to further optimize feature representation and improve prognostic performance.

Beyond micropapillary components, our combined model also identified acinar and papillary components as independent prognostic variables. While pure acinar adenocarcinoma is traditionally associated with favorable outcomes, emerging evidence suggests that even minor admixtures of high-risk patterns (e.g., solid or micropapillary components as low as 5%) can substantially elevate recurrence and mortality risk (26). A predominance of acinar pattern in a micropapillary-containing tumor may indicate a distinct molecular subtype with hybrid aggressive features. Clinical perspective: The key prognostic determinant may not be the absolute percentage of any single pattern, but rather the balance between indolent (acinar) and aggressive (micropapillary/solid) phenotypes. The inclusion of acinar components in our nomogram aligns with these contemporary insights and underscores the model’s ability to capture nuanced histological risk profiles.

The odds ratios for growth patterns as continuous percentage variables require balancing mathematical precision with clinical plausibility. Although the model reports odds ratios per 1% increase, it should be emphasized that this represents a marginal effect. For clinical interpretation, more meaningful increments (e.g., 5% or 10%)—which better reflect real-world pathological assessment—should be considered. While extrapolating the effect to high percentages may seemingly yield very large odds ratios, this is constrained by biological reality: because the percentages of growth patterns sum to 100%, a substantial increase in one component necessarily entails a decrease in others. This interdependence among patterns makes extreme isolated changes clinically implausible. Moreover, the effect magnitudes align with the established prognostic hierarchy (micropapillary > solid > papillary > acinar > lepidic), and the literature supports that even small amounts of high-risk patterns (e.g., 5% micropapillary) significantly impact prognosis. Therefore, we recommend interpreting these effects within the typical clinically observed range (usually 0–30% change for an individual pattern) and caution against over-extrapolation to mathematically possible but biologically unrealistic extremes.

Several limitations of this study warrant acknowledgment. First, its single-center, retrospective design and relatively modest sample size may constrain generalizability. In this cohort, female patients accounted for 60.4%, which is higher than the male-dominated proportion typically observed in the Chinese lung adenocarcinoma population. This may reflect regional variations, selection bias (e.g., a greater tendency among women to undergo early screening), or potential sex-related differences in the distribution of the micropapillary subtype. These possibilities require further validation in multicenter studies to confirm the robustness and clinical applicability of our model. Second, although molecular testing for EGFR, ALK and other gene sites was performed in some patients, the inconsistent testing rates and lack of comprehensive data prevented its inclusion in our predictive model. This represents a significant gap, as these biomarkers are crucial for prognosis and treatment selection in lung adenocarcinoma. Third, the absence of a control group comprising lung adenocarcinoma patients without micropapillary components limits our ability to assess the model’s specificity and the unique prognostic contribution of micropapillary features relative to other subtypes. Fourth, the follow-up period is relatively short (minimum ~1.5 years), which may limit the assessment of long-term survival outcomes. Fifth, the performance gap between the training and test sets, particularly for the pathomics model, suggests potential overfitting. The combined model’s superior performance in the training set did not fully translate to the independent test set, indicating a need for further optimization and validation with larger datasets. Future research should aim to integrate multi-omics data—including genomics, transcriptomics, and digital pathomics—to develop more comprehensive and biologically interpretable prognostic systems.

Conclusions

In summary, this study successfully developed and validated a combined clinicopathological prediction model based on weakly supervised deep learning for stratifying postoperative survival risk in patients with lung adenocarcinoma containing a micropapillary component. The combined model showed promising predictive capability in the training set, though its performance attenuated in independent testing. This work establishes the feasibility of weakly-supervised pathomics for this high-risk subtype but highlights the need for larger validation cohorts before clinical application. This approach provides an objective, scalable tool for identifying high-risk individuals, with the potential to support personalized postoperative management and improve clinical decision-making. Future efforts should focus on external validation across diverse cohorts and the incorporation of multi-omics information to further refine prognostic accuracy and biological relevance.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study has been reviewed and approved by the Ethics Review Committee of the Affiliated Hospital of Jiaxing University (Jiaxing First Hospital) (approval No. 2025-KY-411). All patients have signed the informed consent form.

Data Sharing Statement

Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1-2647/dss

jtd-18-03-234-dss.pdf ^{(103.1KB, pdf)}

DOI: 10.21037/jtd-2025-1-2647