Visceral pleural invasion is a serious prognostic finding in early lung cancer. The visceral pleura is the thin membrane directly covering the lung surface. When a lung tumor grows through this membrane - a finding called visceral pleural invasion (VPI) - it signals more aggressive disease. Studies consistently show that VPI is associated with significantly worse five-year survival rates in early-stage non-small cell lung cancer (NSCLC), even for small tumors under 3 cm.
VPI affects TNM staging and treatment decisions. VPI is formally classified in the TNM staging system as a T descriptor, effectively upstaging tumors from T1 to T2 regardless of size. This distinction is clinically critical: a tumor with VPI may require more extensive surgery, adjuvant chemotherapy, or closer post-operative surveillance compared to an identical tumor without VPI.
VPI is especially important as minimally invasive surgery expands. The growing adoption of segmentectomy - surgically removing only a lobe segment rather than an entire lobe - for small peripheral NSCLC depends on accurate knowledge of lymph node status and pleural involvement. Because VPI correlates strongly with lymph node metastasis (30.8% vs 9.8% in this study's VPI-positive vs negative groups), accurate preoperative VPI prediction is essential for deciding which patients are appropriate candidates for this less extensive surgery.
Conventional CT imaging misses VPI frequently. Current radiological assessment uses CT features like pleural tags, tumor dimensions, and pleural thickness to estimate VPI risk, but these assessments rely on subjective interpretation of two-dimensional image slices. VPI can only be definitively confirmed by pathological examination of resected tissue, creating a clinical dilemma: surgeons must decide on the extent of operation before the definitive answer is available.
556 patients with early-stage lung cancer were studied. This retrospective, single-center study included patients with clinical stage 0 to I NSCLC who underwent complete surgical resection at Tokyo Medical University Hospital between 2011 and 2018. The dataset was split chronologically: 408 patients formed the training cohort (surgery before June 2017) and 148 patients formed the test cohort (surgery from June 2017 onwards), creating a temporal validation that mimics real-world deployment conditions.
The Synapse Vincent AI system analyzed 3D CT images. The AI software used is a beta version of the Synapse Vincent System (Fujifilm Corporation, Japan), a commercially oriented platform designed for automated detection and characterization of pulmonary nodules. It uses a three-dimensional convolutional neural network based on a modified U-Net architecture to segment tumors across the entire lung field in three dimensions, overcoming the limitations of selecting individual 2D CT slices.
The system extracted 22 radiological features with confidence scores. Beyond volumetric measurements (total tumor volume, solid-part volume, ground-glass nodule proportion, and maximum solid length), the AI assigned confidence scores from 0 to 1 for each of 22 qualitative radiological features. These included characteristics like solid nodule, part-solid nodule, ground-glass opacity, pleural contact, pleural indentation, spiculation, lobulated edge, bronchus translucency, and several others that radiologists typically describe subjectively.
VPI was confirmed pathologically after surgical resection. All resected specimens underwent pathological staging including specialized Elastica van Gieson staining, which highlights elastic fibers in the pleura, and D2-40 staining for lymphatic vessels. VPI was defined as tumor invasion beyond the elastic layer of the visceral pleura, consistent with international staging guidelines. This pathological ground truth was used to train and validate the AI prediction model.
Stability selection was used to identify the most robust predictors. Standard variable selection methods like stepwise regression and Lasso (a regularization technique) are sensitive to small changes in the data and can produce inconsistent results across different samples. The stability selection algorithm addresses this by repeatedly applying Lasso to 50 randomly subsampled versions of the training dataset and measuring how consistently each variable is selected. Features chosen in more than 75% of subsamples are considered stable, true predictors.
Three competing prediction models were built and compared. Model 1 used only the 22 AI-derived radiological confidence scores. Model 2 added 3D volumetric parameters (tumor volume, solid-part volume ratios) to those features. Model 3 additionally incorporated patient characteristics like age, sex, smoking history, and histological type. All three models then underwent logistic regression to produce a final prediction formula expressing the probability of VPI for any new patient.
AI confidence scores were transformed before modeling. The original confidence scores ranged from 0 to 1 but tended to cluster near the extremes. A logit transformation was applied after slight adjustment toward 0.5, converting the probability scores into a continuous scale spanning all real numbers. This prevents the statistical distortion caused by ceiling and floor effects when variables pile up at the boundaries of their allowed range.
Performance was measured by AUC, sensitivity, and specificity. Receiver operating characteristic (ROC) analysis generated area under the curve (AUC) values for each model in both training and test cohorts. The optimal probability cutoff was selected in the training cohort at the point where the sum of sensitivity (true positive rate) and specificity (true negative rate) was maximized, then applied unchanged to the test cohort to assess real-world discriminative performance.
Solid nodule was the single strongest predictor of VPI. Among all 22 AI-derived radiological features, 'Solid nodule' demonstrated the highest AUC of 0.787 in the training cohort. This makes intuitive clinical sense: tumors that appear entirely solid on CT (rather than mixed solid-and-ground-glass or pure ground-glass) are more likely to represent more invasive, aggressive cancer capable of penetrating the pleural membrane.
Pleural contact was the second most important feature. The AI feature 'Pleural contact' achieved an AUC of 0.704, reflecting how closely and extensively the tumor surface adheres to the pleural membrane. When both high solid-nodule scores and high pleural-contact scores were present simultaneously, VPI was consistently found pathologically - while high solid score alone or high pleural contact alone was insufficient to predict VPI with confidence.
The final model achieved strong discriminative performance. Model 1, using only 'Solid nodule' and 'Pleural contact,' achieved an AUC of 0.816 in the training cohort and 0.782 in the independent test cohort. At the optimal threshold of 0.343 probability, the model demonstrated sensitivity of 73.9% and specificity of 65.7% in the test cohort. These results indicate the model correctly identifies approximately three-quarters of VPI-positive cases while correctly excluding about two-thirds of VPI-negative cases.
Adding more variables did not improve performance in the test cohort. Counterintuitively, Models 2 and 3 - which included additional 3D volumetric parameters and patient characteristics - outperformed Model 1 slightly in the training cohort (AUC 0.831 vs 0.816) but underperformed in the test cohort (AUC 0.767 vs 0.782). This pattern suggests the simpler two-feature model generalizes better to new patients, and that the more complex models were slightly overfit to the training data.
VPI was found in nearly one-third of patients in this study. Among 556 patients with early-stage NSCLC, 32.2% were found to have VPI after surgery. Of these, the large majority had invasion into but not through the elastic layer (PL1, 23.2%), while a smaller proportion showed deeper invasion through the pleural surface (PL2, 6.7%) or into adjacent structures (PL3, 2.3%). This frequency underscores how common this finding is even in clinical stage 0 to I disease.
VPI was associated with more aggressive tumor characteristics. Patients with VPI were significantly more likely to be male (71.4% vs 44.0%), had larger solid tumor sizes, and were less likely to have adenocarcinoma (70.7% vs 89.5%). Crucially, the rate of lymph node metastasis was more than three times higher in VPI-positive patients (30.8% vs 9.8%), confirming VPI as a marker of systemic metastatic potential even in early-stage disease.
Pathological upstaging affected over one-fifth of patients. Despite being classified as clinical stage 0 to I before surgery based on imaging, 21.8% of patients were ultimately upstaged to pathological stage II through IV after complete resection and examination. VPI is one of the drivers of this upstaging, as it reclassifies T1 tumors to T2 status regardless of tumor diameter - an important consideration for post-operative treatment planning.
The VPI prediction formula uses a simple mathematical equation. The logistic regression model translates the AI confidence scores for 'Solid nodule' and 'Pleural contact' into a VPI probability: Prob(VPI) = 1 / (1 + e raised to the power of 1.0479 minus 0.590 times the logit of solid minus 1.099 times the logit of pleural contact). A probability above 0.343 predicts VPI positive. This transparent formula could be integrated directly into CT reporting software.
3D analysis evaluates the entire tumor volume rather than selected slices. Conventional radiology relies on two-dimensional slices from CT scans, which means the radiologist only sees a subset of the tumor's total relationship with the pleura. AI-driven 3D reconstruction assesses the entire tumor surface and its spatial relationship with the pleura simultaneously, enabling characterization of features like pleural contact extent that cannot be reliably measured from 2D slices alone.
AI provides objective, reproducible measurements independent of observer variation. A key limitation of manual CT interpretation for VPI is interobserver variability: different radiologists, or the same radiologist on different days, may assess the same image differently. The AI system assigns standardized confidence scores using consistent computational algorithms across all patients, eliminating subjective variation and enabling reproducible decision support.
Pleural tags and indentation were not the strongest predictors in this study. Traditionally, pleural tags (linear strands extending from the tumor toward the pleura) and pleural indentation (puckering of the pleural surface) have been considered the most important CT indicators of VPI. However, in this study - which specifically enrolled patients with tumors already in contact with the pleural surface - solid nodule composition and direct contact extent were stronger predictors. This may reflect that pleural tags are most informative when the tumor does not directly touch the pleura.
The study is a first application of AI to 3D CT for VPI prediction. Prior studies have used 2D CT features and some AI analysis on 2D images, but this is the first study to apply AI-driven 3D CT image characterization specifically to VPI prediction. The authors note that 3D analysis enables evaluation of the total solid component volume and its relationship with the pleural surface - metrics unavailable from 2D assessment that may capture the biological risk of pleural invasion more completely.
Single-center retrospective design limits generalizability. The study was conducted at a single Japanese institution over a defined historical period. Patient selection bias is possible, and the patient population (predominantly adenocarcinoma, Japanese patients) may not represent other cancer subtypes or ethnic groups. External validation in diverse multi-center cohorts is needed before this model can be adopted broadly.
Performance declined from training to test cohort. The AUC dropped from 0.816 to 0.782 when moving from the training cohort to the independent test cohort. Part of this decline is attributable to performance of the 'Solid nodule' feature being lower in the test cohort (AUC 0.741 vs 0.786 in training). This highlights the importance of temporal validation rather than cross-validation alone, as patient populations and imaging protocols can shift over time.
No direct spatial correlation between CT findings and pathology was established. The study demonstrated statistical association between AI-derived CT features and pathologically confirmed VPI, but did not validate that the CT feature of pleural contact corresponds to the exact anatomical location where VPI was found histologically. Such spatial correspondence studies would strengthen confidence in the biological interpretability of the AI predictions.
The AI system remains a research beta version not yet commercially available. The Synapse Vincent AI software used in this study is an experimental research tool under development, not a commercially released product. Translating these findings into clinical practice will require regulatory validation, prospective testing, and integration with existing radiology and surgical workflow systems. Nevertheless, the results establish a strong proof of concept for AI-assisted VPI prediction in routine preoperative CT evaluation.