High-Grade Patterns in Lung Adenocarcinoma: Lung adenocarcinoma (LUAD) is the most common type of lung cancer, and its prognosis depends heavily on its histological grade. High-grade patterns including micropapillary and solid subtypes are associated with significantly worse outcomes, including higher recurrence and lower survival rates. Identifying these patterns before surgery guides treatment planning.
Current Diagnostic Limitations: Definitive grading currently requires surgical resection and pathological examination, making preoperative prediction challenging. While CT imaging can hint at tumor characteristics, it cannot reliably distinguish high-grade from low-grade tumors. PET/CT adds metabolic information through fluorodeoxyglucose (18F-FDG) uptake but has not been fully integrated with deep learning approaches.
Study Approach: This study from a Chinese hospital enrolled 303 LUAD patients who underwent preoperative 18F-FDG PET/CT and subsequent surgery. Researchers developed deep learning models from CT and PET images and combined them with clinical and metabolic parameters into a nomogram to predict high-grade patterns non-invasively.
Dual Strategy - Deep Learning and Nomogram: The study pursued two complementary predictions tools: image-based deep learning (DL) models that directly analyze CT or PET scan pixels, and a clinical-metabolic (CM) nomogram combining interpretable clinical variables. Both approaches were compared individually and in combination.
Patient Cohort and Data: 303 LUAD patients were retrospectively analyzed, split into training and validation cohorts. All patients had preoperative 18F-FDG PET/CT imaging and confirmed pathological diagnoses post-surgery. High-grade patterns (micropapillary or solid components) were identified by pathologists as the study endpoint.
ResNet-18 Deep Learning Models: ResNet-18, a well-validated 18-layer residual convolutional neural network, was applied to tumor regions of interest extracted from CT images (CT-DL model) and from PET images (PET-DL model). A combined PET/CT-DL fusion model was also constructed. Tumor regions were manually delineated by radiologists to provide the input to each model.
Clinical-Metabolic Nomogram Variables: Logistic regression with LASSO regularization identified the most predictive clinical and metabolic variables from the PET/CT reports. The final CM nomogram incorporated four variables: clinical stage, serum CYFRA21-1 (a lung cancer marker), SUVmean (mean standardized uptake value on PET), TLG (total lesion glycolysis), and tumor size.
Model Evaluation: Area under the ROC curve (AUC) was the primary metric. Calibration curves assessed agreement between predicted probabilities and actual outcomes. Decision curve analysis (DCA) evaluated net clinical benefit. The DeLong method compared AUC values between models.
CT Deep Learning Leads Performance: The CT-DL model achieved the best individual performance with an AUC of 0.895 in the validation cohort, outperforming the PET-DL model (AUC 0.817) and the PET/CT combined model (AUC 0.850). This suggests CT texture features captured by the neural network are more informative than PET metabolic patterns alone for grading.
CM Nomogram Performance: The clinical-metabolic nomogram achieved an AUC ranging from 0.817 to 0.977 across different validation subsets, with the combined model reaching up to 0.977. The nomogram incorporating all four variables provided good calibration with the calibration curves closely following the ideal diagonal line.
Combined Model Advantage: When the CT-DL model scores were incorporated into the CM nomogram, the combined model achieved the highest AUC, demonstrating that imaging-derived deep learning features complement and enhance interpretable clinical variables. Decision curve analysis confirmed net clinical benefit across a wide probability threshold range.
Metabolic Correlates of High-Grade Patterns: Higher SUVmean and TLG values were significantly associated with high-grade LUAD, confirming that more metabolically active tumors on PET tend to harbor aggressive histological features. CYFRA21-1 serum levels also correlated significantly with high-grade patterns, adding a blood biomarker component.
Surgical Planning Impact: Preoperative identification of high-grade patterns can guide the extent of lung resection. Patients with predicted high-grade tumors may benefit from more aggressive resections (lobectomy rather than sublobar resection) and more extensive lymph node sampling to ensure complete tumor removal.
Adjuvant Therapy Guidance: High-grade LUAD has higher recurrence rates and may benefit from adjuvant chemotherapy or targeted therapy (in EGFR/ALK-positive cases). Non-invasive preoperative prediction allows oncologists to discuss adjuvant treatment plans with patients before surgery, facilitating shared decision-making.
Avoiding Repeat Biopsies: The nomogram and DL models could reduce reliance on repeat biopsies for grading, which carry procedural risks including pneumothorax and bleeding. Routine pre-surgical PET/CT is already standard in many centers, meaning no additional imaging would be required.
CYFRA21-1 as a Practical Marker: The inclusion of serum CYFRA21-1 in the nomogram is clinically important because it is a widely available, inexpensive blood test. Its independent predictive value for high-grade patterns adds a simple, accessible component to the preoperative assessment toolkit.
Retrospective Design Limitations: This study analyzed retrospectively collected data from a single center in China. The patient population, imaging protocols, and pathology practices may differ from institutions in other countries, limiting direct generalizability of the AUC values.
Prospective Multicenter Validation Needed: External validation across multiple institutions with diverse patient demographics, PET/CT scanner types, and imaging protocols is essential before these models can be recommended for clinical use. Prospective studies would also allow assessment of whether model-guided decisions actually improve patient outcomes.
Integrating Molecular Markers: Future work could incorporate EGFR/ALK mutation status, KRAS mutations, or other molecular features into the nomogram, potentially improving prediction for specific patient subgroups. The interaction between metabolic PET parameters and driver mutations in predicting grade warrants investigation.
Automated Region Segmentation: Currently, tumor regions are manually delineated by radiologists before feeding into the DL model. Automating this segmentation step using AI-based tools would be necessary for scalable clinical deployment and would reduce inter-observer variability in the input data.