Lung adenocarcinoma is the most prevalent form of lung cancer, accounting for approximately 50 percent of all diagnoses. It exhibits pronounced heterogeneity because tumors commonly contain mixtures of multiple growth patterns coexisting within the same specimen, making accurate subtype identification critical for treatment planning and prognosis.
Different histologic subtypes carry markedly different clinical implications. The lepidic growth pattern is considered the least invasive, while micropapillary and solid patterns are the most invasive and are linked to higher recurrence rates and poorer survival. Acinar, papillary, and cribriform patterns fall in between, with acinar and cribriform patterns recently merged under the 2021 WHO classification update.
Microscopic examination of tissue slides is labor-intensive, time-consuming, and prone to inter-observer variability. In challenging cases such as atypical breast cancer, diagnostic concordance between pathologists can be as low as 48 percent, illustrating how even expert clinicians can disagree on complex histological interpretations.
Deep learning and computational pathology offer a path to more consistent, scalable subtype classification, but these approaches require large, high-quality, well-annotated datasets. Existing public datasets for lung adenocarcinoma are limited and do not reflect the demographic profile of Chinese patients, who account for the largest absolute number of lung cancer cases worldwide.
The Chinese Lung Adenocarcinoma WSI Dataset (CLWD) was assembled retrospectively from 210 patients seen between 2020 and 2023. A total of 408 hematoxylin and eosin stained whole-slide images were collected, digitized, and annotated at the First People's Hospital of Yunnan Province, covering seven subtypes: in situ, acinar, cribriform, lepidic, micropapillary, papillary, and solid patterns.
Slides were scanned using an SQS-600P scanner at 80 times magnification, yielding a resolution of 0.1 micrometers per pixel and a level of detail comparable to direct microscopic observation. When an initially available slide was suboptimal for digitization, representative sections were recut from formalin-fixed paraffin-embedded tissue blocks and restained before scanning.
Subtype labels were assigned by board-certified lung cancer pathologists following established predominant-pattern criteria from both the 2015 and 2021 WHO classification systems. The dataset provides annotation layers compatible with both systems, allowing researchers to benchmark under either the 7-class or 6-class schema. Immunohistochemistry was used when clinically indicated to support diagnosis, and diagnostically ambiguous cases were excluded during quality control.
Comprehensive clinical metadata accompanies each slide, including patient sex, age at diagnosis, pathological diagnosis category (adenocarcinoma in situ, minimally invasive adenocarcinoma, or fully invasive adenocarcinoma), and detailed subtype labels. The dataset is publicly available via Figshare and a pathology image repository, with a JPEG version also hosted on Hugging Face for convenient access.
Whole-slide images are too large to process as single images, requiring a specialized approach called multiple instance learning. In this framework, each slide is treated as a bag of smaller image patches, and the model learns slide-level labels from the aggregated features of these patches without needing precise annotations at the pixel or patch level, transforming the problem into a weakly supervised task.
Three established classification architectures were benchmarked on the dataset. CLAM uses attention-based deep learning to automatically identify diagnostically relevant subregions within whole-slide images and incorporates instance-level clustering to refine feature representations and improve interpretability. It is particularly well suited for multi-class subtyping under weak supervision.
TransMIL applies the self-attention mechanism of Transformer neural networks to capture long-range dependencies across image patches, modeling both morphological and spatial relationships between tissue regions. This approach addresses a key limitation of conventional multiple instance learning methods that treat patches independently and ignore structural relationships.
The Graph Transformer combines graph neural networks with Transformer mechanisms, constructing a relational graph among image patches to model their spatial and semantic interactions. Adaptive attention weights are assigned to each patch node, enabling weighted aggregation into slide-level representations that reflect tissue architecture.
All models were implemented in Python using PyTorch 2.3.0 on a server with four Nvidia V100 GPUs. Patient-level data splitting was enforced to prevent data leakage, with a 9:1 or 8:1:1 split for training and testing, and five-fold cross-validation used within the training set for robust evaluation.
All three models achieved strong discrimination for in situ adenocarcinoma on the CLWD dataset. Area under the receiver operating characteristic curve values for in situ classification ranged from 88.08 to 96.16 across CLAM, Graph Transformer, and TransMIL, reflecting the morphologically distinct and recognizable appearance of this non-invasive pattern.
CLWD consistently yielded stronger and more stable discrimination than the external Dartmouth DHMC dataset across most subtypes and models. Where the DHMC cohort showed larger class-dependent variability and reduced separability for certain patterns such as papillary and acinar under some models, CLWD maintained more consistent performance, supporting its utility as a high-quality benchmark resource.
Performance was comparatively lower for papillary and acinar subtypes, consistent with the known morphological overlap among intermediate-risk growth patterns. The minimally invasive adenocarcinoma category was the most challenging to separate in the three-class invasion-status task, partially attributable to its smaller sample size in the dataset.
Results were consistent across two different data splitting strategies, a 9:1 patient-level split with internal cross-validation and a stratified 8:1:1 fixed split. The stability across these experimental designs supports the learnability and reliability of the CLWD annotations as a training resource for computational pathology model development.
Several limitations should be considered when using the CLWD dataset. All images were acquired at a single institution using a uniform scanning platform, which may limit the generalizability of models trained on this data to slides prepared at institutions with different staining protocols or scanning equipment.
Class imbalance among subtypes is a practical challenge. In situ adenocarcinoma has 80 slides, papillary has 84, and micropapillary has 64, while cribriform has only 31 and lepidic has only 30, creating potential biases in model training. Individual whole-slide images may also contain multiple histological subtypes simultaneously, further complicating classification.
The absence of immunohistochemistry marker stained slides limits the dataset's utility for analyses involving molecular biomarkers or multi-stain approaches. Formal inter-rater reliability metrics such as Cohen's kappa across the full dataset were not computed in the current release, which the authors acknowledge as a limitation given that subtype annotation involves inherent inter-observer variability.
Despite these limitations, the dataset supports a wide range of research applications including machine learning model development and validation, digital pathology educational applications, and integration with other publicly available resources. The public availability of both the raw SVS images and JPEG versions, along with the accompanying code on GitHub, maximizes accessibility for the global research community.
The CLWD dataset addresses a critical gap in diversity among existing public lung adenocarcinoma pathology resources. Because China has the largest absolute number of lung cancer cases globally due to its population size, a dataset specifically curated from Chinese patients is essential for developing AI models that generalize across different population demographics.
The ability to accurately identify high-risk early-stage subtypes such as acinar, micropapillary, and solid patterns is directly relevant to clinical decisions about adjuvant therapy and surveillance intensity. Computational tools trained on CLWD could support more consistent subtype identification that informs these critical treatment decisions.
The inclusion of WHO 2021 compatible annotation layers alongside the older 2015 classification schema allows researchers to develop and benchmark models under both systems as the field transitions between classification standards. This dual annotation approach increases the dataset's long-term utility across research conducted under different guideline frameworks.
The dataset's public availability via multiple platforms, including Figshare, a dedicated Pathology Image Repository, and Hugging Face, along with open-source code on GitHub, is designed to enable broad international collaboration and lower barriers for computational pathology researchers without extensive data collection infrastructure of their own.