Our findings strongly indicate that integrating radiological CT features with radiomics models can significantly enhance the ability to differentiate between patients with IMA and those with invasive non-mucinous LADC, surpassing the diagnostic efficacy achieved using only radiological CT features or radiomics. Therefore, we successfully developed a combined nomogram with excellent discrimination and calibration to accurately distinguish between patients with SPN-type IMA and those with invasive non-mucinous LADC.
SPN-type IMA is the primary manifestation of IMA, and ground-glass nodules are rare [17]. This manifestation distinction may be attributed to IMA’s histological characteristics, such as mucin-rich tumor cells, fibrosis, and central fibrosis with alveolar mucin-filled spaces [6]. Recent studies have indicated a higher prevalence of SPN-type IMA compared with pneumonia-type IMA [8, 15]. Therefore, differentiating between SPN-type IMA and non-mucinous LADC is crucial, particularly in the early stages. This study found that SPN-type IMA exhibited more vascular convergence signs, pleural indentation, and advanced clinical stage than invasive non-mucinous LADC. These observations may be attributed to mucus leakage from the nodule's margin and the migration of macrophages along the alveolar walls and pores, leading to signs of vascular convergence and pleural indentation. These findings are consistent with those of Wu et al., Zhang et al., and Cha et al. [16–18]. However, the clinical prediction model using CT radiological features demonstrated a modest diagnostic efficiency, as evidenced by the AUC values of 0.714 and 0.715 for the training and test sets, respectively. This could be due to the subjective interpretation of the radiological features, which may lack consistency, typicality, and generalizability, particularly during the early stages.
Radiomics is an advanced, noninvasive approach that uses intelligent algorithms to construct models based on original medical images. This method yields additional insights and potentially reveals pertinent phenotypic information by capturing tumor heterogeneity [19]. To our knowledge, no study has investigated the use of radiomic models in distinguishing between clinical stage IA SPN-type IMA and non-mucinous LADC. Notably, our study revealed an independent association between higher rad scores and IMA in the radiomics model. This observation can be explained by the heightened IMA heterogeneity, attributed to the abundant presence of mucin within the cytoplasm of IMA tumor cells and the aggressive nature of tumor neovascularization, closely associated with its malignant behavior [17]. The radiomics model achieved AUC values of 0.770 and 0.753 for the training and test sets, respectively, indicating higher diagnostic efficiency than that of the clinical model. This is because radiomics provides a notable advantage in evaluating tumor imaging phenotypes. It enables the extraction of quantitative features from CT images, allowing the identification of additional factors that cannot be easily comprehended visually using CT radiology. In addition, through five cross-validated analyses, the radiomics model demonstrated favorable predictive performance, with a mean AUC value of 0.771 in differentiating IMA from invasive non-mucinous LADC, indicating the model’s good robustness and avoidance of overfitting.
By amalgamating crucial data from clinical and radiomics models, the combined model accurately distinguished between IMA and invasive non-mucinous LADC. It performed better than the clinical or radiomics models when used independently. We hypothesized that integrating diverse and abundant information within the combined model would increase diagnostic efficiency when distinguishing IMA from invasive non-mucinous LADC.
This study has some limitations. First, it is crucial to acknowledge our study’s retrospective nature and the relatively low incidence rate of IMA stemming from the rarity of this condition. In addition, potential selection bias in our research arose from including only patients with postoperative pathologic results. Second, the nature of our bicentric study introduced variations in acquisition parameters, image quality, and potential co-registration errors, which could confound the analysis by contributing to the sources of variability. Furthermore, artifacts and technical limitations may affect the reliability and reproducibility of radiomic features. Third, we could not develop a predictive model for disease outcomes due to the relatively short postoperative follow-up period. Finally, our study focused solely on SPN-type IMA and invasive non-mucinous LADC with a diameter of ≤ 3 cm, which may limit the generalizability of our findings to other stages.
In conclusion, our study successfully developed an innovative model using preoperative radiological and radiomic features to effectively differentiate between IMA and invasive non-mucinous LADC. We transformed this combined model into a nomogram that accurately quantified the risk of IMA. Notably, the nomogram exhibited exceptional discrimination and calibration, highlighting its potential value in clinical practice.