Identification of non-invasive PET-CT features for differentiation and prognostication of lung adenocarcinoma and squamous cell carcinoma using radiomics and foundation model

This study demonstrates the potential of non-invasive conventional radiomics and advanced deep learning-based foundation models (DINOv2) for differentiating between ADC and SCC and for identifying prognostic subgroups. Features were extracted from pretreatment PET-CT images of patients diagnosed with ADC and SCC. The performance of features in distinguishing between ADC and SCC was evaluated using statistical tests. Further, patients were clustered based on these features, and Kaplan-Meier survival analysis was performed to assess the prognostic value of the identified clusters. The results demonstrated that both conventional radiomics and DINOv2 features could effectively differentiate between ADC and SCC. First-order radiomic features, including Entropy, Uniformity, and Mean intensity from CT images, and Median and Mean intensity from PET images, showed statistically significant differences between the two cancer types (p < 0.05). Notably, DINOv2 features exhibited consistently higher statistical significance (p < 0.000001) in distinguishing between ADC and SCC, highlighting the superior discriminatory power of deep DINOv2 features. Survival analysis revealed that clustering patients based on both radiomics and DINOv2 features could identify subgroups with distinct survival outcomes. Further validation in larger, multi-center cohorts is warranted to confirm these findings and translate them into clinical practice for improved diagnosis and personalized treatment strategies.