Lung cancer is the deadliest cancer worldwide. With approximately 1.8 million deaths and 2.5 million new cases in 2022 alone, lung cancer accounts for 18.7% of all cancer deaths. Non-small cell lung cancer makes up 85% of cases, including adenocarcinoma and squamous cell carcinoma subtypes with distinct morphological and molecular profiles requiring accurate classification for proper treatment.
Histopathology is the gold standard but has limitations. While CT, MRI, and PET scans provide useful structural and metabolic information, they cannot definitively categorize tumor subtypes. Histopathological examination of biopsied tissue provides cellular and subcellular resolution, enabling pathologists to identify key morphological characteristics such as architectural patterns, mitotic activity, and nuclear features. However, manual analysis is labor-intensive, subjective, and prone to interobserver variability.
Deep learning is transforming digital pathology. Convolutional neural networks (CNNs) have demonstrated outstanding performance in image classification tasks and can automatically learn hierarchical representations from raw tissue images without manual feature engineering. However, most existing approaches use a single deep learning framework, missing opportunities to capture complementary feature information from different representation learning paradigms.
The gap this work addresses. While graph neural networks and self-supervised learning (SSL) approaches have individually shown promise in medical imaging, their combination with optimized feature selection for histopathological lung cancer classification remains largely unexplored. Most prior work focuses on a single SSL framework or does not systematically remove redundant features from high-dimensional SSL representations before classification.
Self-supervised learning learns without labeled data. SSL methods learn rich feature representations from unlabeled data, which is abundant and cheap to obtain, then transfer those representations to downstream classification tasks. This is especially valuable in medical imaging where labeled datasets are small and expensive to annotate. The researchers combined three fundamentally different SSL approaches to capture complementary aspects of tissue morphology.
SimCLR: contrastive learning from augmented views. SimCLR generates two augmented versions of each image and trains the model to maximize agreement between these paired views while pushing apart representations from different images. This approach emphasizes fine-grained local pattern discrimination and learns instance-level invariance, helping the model recognize the same tissue structure across different orientations or stains.
DeepCluster: semantic grouping through clustering. DeepCluster alternates between clustering extracted feature vectors using k-means and using the resulting cluster assignments as pseudo-labels to update the CNN. This iterative process teaches the model to discover global structural patterns and inter-class semantic groupings without any human annotation, capturing coarse-grained distinctions between tissue classes.
BYOL: bootstrap learning without negative pairs. Bootstrap Your Own Latent (BYOL) uses two networks, an online and a target network, where the online network learns to predict the target network's representations of augmented views. The target network is updated via exponential moving average rather than gradient descent, providing stable training without requiring negative sample pairs. BYOL captures consistent, instance-level morphological features robust to stain and acquisition variations.
Custom CNN extracts 128 base features per image. A custom CNN processes each 224x224 pixel histopathological image through stacked convolutional layers with batch normalization and ReLU activation, followed by max pooling. This produces 128 features per image that capture spatial hierarchies of tissue morphology. These features are then independently optimized by each of the three SSL methods, each producing its own 128-dimensional enhanced representation.
Feature fusion creates 384-dimensional representations. The 128-dimensional outputs from SimCLR, DeepCluster, and BYOL are concatenated to create a 384-dimensional fused feature vector. This fusion brings together complementary information: SimCLR's contrastive alignment highlights discriminative local patterns, DeepCluster's clustering adds semantic grouping information, and BYOL's bootstrap learning contributes stable cross-augmentation invariance. The combination creates richer and more diverse representations than any single method alone.
mRMR selects 100 non-redundant features. The 384 fused features are then filtered using Minimum Redundancy Maximum Relevance (mRMR), a filter-based feature selection method that simultaneously maximizes relevance to the target class and minimizes redundancy among selected features. mRMR selects the top 100 features from the 384, reducing dimensionality while retaining the most discriminative information. This step is critical because raw fused SSL representations may contain redundant or noisy dimensions.
Vision GCN performs graph-based classification. The 100 selected features are used to build a graph where each sample (image) is a node connected to its top-15 most similar samples by cosine similarity. The Vision Graph Convolutional Network (VGCN) then applies multi-head graph attention to aggregate information across neighboring nodes, capturing structural and spatial relationships between tissue samples. Focal Loss addresses class imbalance, and dropout regularization prevents overfitting.
Three publicly available histopathology datasets used. The primary dataset was LungHist700, containing 691 high-resolution histopathology images (1200 x 1600 pixels) from 45 patients at 20x and 40x magnification, representing three classes: 280 adenocarcinoma, 260 squamous cell carcinoma, and 151 normal tissue. Additional validation was performed on LC25000 (250 images per class from 86 whole-slide images) and TCGA-UT (500 patches each from adenocarcinoma and squamous cell carcinoma classes).
Patient-level data splitting prevents leakage. All datasets were divided into 70% training, 15% validation, and 15% testing subsets using stratified splitting. Critically, data splitting was performed at the patient level, meaning all patches from the same patient were restricted to a single partition. This prevents data leakage between training and test sets that would artificially inflate performance estimates.
Balanced accuracy used as primary metric. Because the datasets are class-imbalanced (particularly LungHist700 with only 151 normal vs. 280 adenocarcinoma images), balanced accuracy (BAC) was used as the primary metric. BAC assigns equal weight to each class regardless of sample count, providing a more honest assessment than standard accuracy for imbalanced classification problems. Focal Loss was used during training to address class imbalance further.
Systematic hyperparameter optimization conducted. Grid search was applied to optimize key hyperparameters including SimCLR temperature (0.5), DeepCluster number of clusters (10), BYOL momentum (0.99), VGCN dropout (0.4), learning rate (0.003), and top-K neighbors (15) for graph construction. All tuning was conducted exclusively on the validation set, with the test set held out entirely for final evaluation.
Proposed method achieves 97% balanced accuracy. The full MuSSL+mRMR+VGCN pipeline achieved a balanced accuracy of 0.97 on the LungHist700 test set, compared to 0.86 for the plain VGCN baseline. Precision, recall, and F1-score all reached 0.97 in cross-validation, with a standard deviation of only 0.01 across five folds, indicating stable and consistent performance.
Each component adds measurable value. Individual SSL methods applied alone (SimCLR+VGCN: 0.77, DeepCluster+VGCN: 0.78, BYOL+VGCN: 0.85) actually reduced performance compared to plain VGCN (0.86), showing that individual SSL can introduce noise. Fusing all three methods (MuSSL+VGCN: 0.94) boosted performance substantially. Adding mRMR feature selection (MuSSL+mRMR+VGCN: 0.97) provided the final performance jump.
mRMR outperforms LASSO for feature selection. LASSO feature selection (MuSSL+Lasso+VGCN) achieved only 0.94 BAC, the same as fusion without selection, because LASSO's L1 regularization can eliminate weakly weighted but relevant features. mRMR's explicit optimization for both relevance and non-redundancy was more effective, retaining a more informative and diverse feature subset that better served VGCN's graph-based learning.
t-SNE visualization confirms improved class separability. T-distributed Stochastic Neighbor Embedding (t-SNE) plots comparing original features versus MuSSL+mRMR features show that the proposed pipeline dramatically reduces class overlap in feature space. The three tissue classes (normal, adenocarcinoma, squamous cell carcinoma) form clearer, more distant clusters after feature optimization, confirming that the pipeline is genuinely improving the discriminative structure of the learned representations.
Near-perfect accuracy on two additional datasets. On the LC25000 dataset, MuSSL+mRMR+VGCN achieved a BAC of 1.00 with precision, recall, and F1-score of 0.99, representing a 5.2% improvement over plain VGCN (0.95). On the TCGA-UT dataset, the method achieved BAC of 0.99 with all other metrics at 0.99, a 4.2% improvement. These consistent gains across three datasets from different institutions and acquisition protocols support the framework's generalizability.
Squamous cell carcinoma classified more accurately than adenocarcinoma. The F1-score for squamous cell carcinoma was 0.94 versus 0.88 for adenocarcinoma on the primary dataset. This difference reflects the histopathological characteristics of each subtype: squamous cell carcinoma presents with well-defined keratinization, polygonal morphology, and strong intercellular junctions that produce clear structural boundaries aligned with graph-based modeling. Adenocarcinoma is more heterogeneous, with irregular nuclear morphology and variable glandular architecture that creates greater intra-class variability.
Statistical significance confirmed. Paired t-tests were conducted to confirm that performance differences were not due to random variation. Plain VGCN achieved a p-value of 0.000046 versus the proposed method, meaning its lower performance is highly statistically significant. Most individual SSL methods also showed p-values well below 0.05, confirming genuine improvements from the full MuSSL+mRMR+VGCN pipeline.
Computation time is manageable. The full pipeline runs in approximately 7 minutes and 41 seconds, compared to 1 minute 18 seconds for plain VGCN. The custom CNN feature extraction step requires approximately 2 hours and 11 minutes during initial training, but this is a one-time cost. The additional computation time of the proposed method is a reasonable trade-off given the substantial accuracy gains it provides.
Dataset diversity limits real-world generalizability. All three datasets used in the study have specific biases. LungHist700 has limited institutional diversity, LC25000 uses patch-based sampling that may inflate benchmark performance, and TCGA-UT has domain shift challenges from variable slide preparation protocols. These limitations mean the model may not perform identically when applied to slides from different hospitals using different scanners or staining procedures.
Computational complexity is a deployment barrier. The multi-SSL fusion pipeline is computationally intensive compared to single-model approaches. For real-time clinical deployment, this may require the development of lightweight alternatives such as knowledge distillation or model pruning to compress the pipeline into a more efficient form without sacrificing the representational richness gained from multi-SSL fusion.
Interpretability needs improvement for clinical acceptance. While the VGCN classifier captures spatial relationships between samples, the pipeline does not currently provide pathologists with easy-to-understand explanations of which tissue features drive its predictions. Linking selected mRMR features to recognized histopathological patterns would improve clinical trust and allow pathologists to verify or override model decisions when needed.
A clear roadmap to clinical use. The authors outline four steps for clinical translation: rigorous validation on diverse multi-institutional prospective datasets, integration into radiology and pathology workflows with explainability mechanisms, regulatory approval with evidence of safety and demographic fairness, and robust technical deployment with ongoing domain shift monitoring. Future work should also explore semi-supervised learning, multi-modal data integration, and stain normalization to improve cross-domain robustness.
A new state-of-the-art for histopathology classification. The MuSSL+mRMR+VGCN framework achieves 97% balanced accuracy on the primary LungHist700 dataset, outperforming prior methods including ensemble CNN with reinforcement learning (92%), graph-based sparse PCA (92%), deep CNN on histology slides (91%), and hybrid SVM+CNN models (94%). This represents a meaningful advance in automated lung cancer subtype classification from histopathological images.
The power lies in principled combination. The key innovation is not any single component but the principled combination: SimCLR contributes contrastive local feature alignment, DeepCluster adds semantic clustering structure, and BYOL provides stable instance-level invariance. Together, they create a feature space richer than any single SSL approach. mRMR then distills this richness into the most discriminative non-redundant subset, and VGCN exploits inter-sample structural relationships for final classification.
Individual SSL is insufficient; fusion with selection is essential. A critical and counterintuitive finding is that individual SSL methods applied alone reduced performance below plain VGCN. This demonstrates that single-framework feature enhancement can introduce noise or remove information present in the original CNN features. Only when multiple SSL representations are fused and then filtered by mRMR does the approach reliably surpass the baseline across all datasets.
Immediate value as a decision support tool. Even in its current form, the framework could assist pathologists by providing rapid, consistent pre-screening of histopathological slides, flagging uncertain cases for expert review, and improving throughput in high-volume diagnostic settings. With further validation and interpretability improvements, it has the potential to become a reliable adjunct tool for lung cancer subtype classification in clinical practice.