With widespread use of high-resolution CT (HRCT), lung cancer is increasingly being detected at an early stage and small peripheral lung cancers are increasingly treated with surgical resection. Adenocarcinoma (ADC) is the most common lung cancer type. According to the World Health Organization classification of lung ADC (1), invasive adenocarcinoma (IAC) is divided into five pathological types, namely lepidic, acinar, papillary, micropapillary, and solid subtypes. The micropapillary pattern (MPP) is a marker for poor outcome (2–6). Lee (7) found that the overall survival and disease-free survival were worse in cases with minimal areas of MPP (accounting for 1-5% of the entire tumor). Since this adverse prognostic effect, identification of MPP may potentially be instructive for surgical plan and further aggressive adjuvant treatment.

Tumor size and MPP are important prognostic factors for surgical outcome in patients with early stage lung ADC. The use of limited resection has gradually increased for patients with non-small cell lung cancer with size ≤ 2 cm. Notably, MPP accounting for more than 5% of the entire tumor is an independent risk factor for recurrence and poor outcome of lung ADC with size ≤ 2 cm (3, 8, 9), suggesting that limited resection may not be the optimal surgical approach for such patients. Thus, preoperative confirmation of MPP (constituting> 5%) in ADC with size ≤ 2 cm has importance for selection of surgical procedure. Due to technical reasons, preoperative histological examination cannot be performed for many peripheral small tumors.

Radiomics can objectively and quantitatively analyze imaging features that reflect tumor heterogeneity. Previous studies showed that radiomics can be used to preoperatively detect ADC with MPP or MPP/solid pattern (10–18). However, previous reports of predicting MPP subtypes using radiomic analyses are with some limitations. First, previous studies have been conducted on the predictor MPP for T1-stage ADC (17, 18), including patients of tumors with size ≥ 2 cm. Because the preferred surgical procedures for ADC with sizes ≤ 2 cm or > 2 cm are different, prediction of MPP in ADC with size ≤ 2 cm can aid in determining the optimal resection method. In addition, previous studies considered only tumor size and overlooked different imaging characteristics between ground glass opacity (GGO) and solid nodule. These radiomics studies combined solid and GGO nodules but the unequal number of these two nodule types may have introduced bias in these study results. Furthermore, these studies were performed on small populations and lacked external validation.

Therefore, the purpose of our retrospective study was to develop an optimal radiomics model for preoperative prediction of MPP with ADC of size ≤ 2 cm. First, we included a large number of ADC patients with peripheral tumor size ≤ 2 cm from three institutions. Second, nodule type was used for stratification and it was combined with HRCT radiomics characteristics to construct two models. Model 1 included all nodule types (solid and GGO). Because most ADC with MPP of size ≤ 2 cm were solid (19), the GGO type was excluded for model 2; only solid nodules were used to construct model. Third, independent external validation was used to validate the two model’s generalization ability.

In this multicenter study, stratification was based on nodule types to construct two radiomics models for preoperative prediction the MPP in peripheral lung ADC with size ≤ 2 cm. The AUC of model 1 (which included solid nodules and GGO) and model 2 (solid nodules only) were 0.91 and 0.78, respectively. Both models had good prediction ability but the prediction performance of model 2 was lower compared to that of model 1, suggesting that GGO affects prediction model performance. In the external validation set, the AUCs for the two models were 0.82 and 0.72, respectively. The DeLong test suggested a difference in AUC between the training and validation sets for model 1, showing poor generalizability. However, no significant difference was noted in AUC between the training and validation sets in model 2, showing good generalizability. Therefore, radiomics can be used as a convenient and non-invasive biomarker for preoperative prediction MPP in peripheral lung ADC with size ≤ 2 cm and to guide the diagnosis and treatment.

Previous studies proved that tumor size is an independent factor for postoperative prognosis in non-small cell lung cancer. Additionally, lung ADC with size 2.1–3 cm has a significantly higher local recurrence rate than ADC < 2 cm (21). Tumor size also affects the choice of surgical procedure for T1-stage peripheral lung ADC. Su (22) found that in ADC patients with tumor size ≤ 2 cm and MPP < 0–5%, limited resection is suitable (3, 8), whereas lobectomy may be required for patients with MPP > 5% (9). Therefore, prediction MPP in lung ADC with size ≤ 2 cm can guide surgical resection strategy.

Many previous studies have showed that CT image-based radiomics analysis can be used to predict ADC with MPP with a prediction performance of 0.7–0.98 (11–18). However, there are inconsistencies and significant differences among the results of previous studies. MPP and solid pattern appear mostly solid on CT, while solid, mGGO, and pGGO types are present in other ADC subtypes. Radiomics identification of GGO and solid lesions is simple. We speculate that the ratio of GGO and solid nodules in ADC without MPP dataset may affected the performance of the prediction model. Many previous studies using radiomics analysis combined GGO and solid nodules in the non-MPP dataset, and the proportion of GGOs was unknown. Chen (11) achieved a prediction performance of 0.86, but GGO and solid nodules accounted for 78% and 22%, respectively, in the ADC without MPP group. Park (12) obtained a prediction performance of 0.98 but the subtypes in non-MPP dataset only consisted of lepidic subtype. The prediction performance decreased to 0.84 when the dataset consisted of papillary and acinar subtype. The dataset by Wang (13) consisted of entire pGGO type for predictor MPP in ADC and the performance was below 0.80. In the current study, the AUC for the prediction performance of model 1 (solid nodules and GGO) was 0.91, but AUC decreased to 0.78 in model 2 after excluding GGO type (69%). In the final external validation, the AUC of model 1 was significantly decreased, may be due to the different ratio of GGO in the training and validation sets. Importantly, the proportion of GGO was lower in the validation set than the training set. These results suggest that the GGO type in the non-MPP dataset contributed to bias in the study results and may explain the differences in results from previous studies.

Notably, there are significant differences in the characteristics of the two prediction models. First, in model 1, the most important characteristic was mean, which was not included in model 2. Mean is a first-order characteristic related to the CT density of the lesion. The CT value of solid nodules is significantly higher than of GGO. Second, first-order characteristic parameters accounted for around half of the parameters in model 1. This may be because GGO is rare in ADC with MPP, resulting in higher contribution to first-order parameters in the histogram. In contrast, there were fewer first-order characteristics in model 2, which means that the role of density characteristics in non-solid nodule prediction is not significant. The prediction of model 1 relied more on texture features and other high-order parameters; therefore, the prediction performance of model 2 was lower than that of model 1. This emphasizes that nodule type is an important factor affecting the prediction MPP in early stage ADC, as well as the necessity of stratifying nodule type when constructing models to predict ADC with MPP. However, the proportion of solid nodules was significantly higher than GGOs (pGGO and mGGO) in ADC with MPP of size ≤ 2 cm (19). Therefore, preoperative prediction of ADC with MPP for solid nodules with 2 cm or less has great clinical value for guiding surgical treatment.

This was a multicenter study and external validation was performed for both models. The results were significantly different between the models, and the generalizability of model 2 was validated in external validation. Although the AUC of model 1 was 0.91 in the training set, which was significantly lower in external validation. There were statistically significant differences in AUC between the training set and external validation set, thereby raising concern over its generalizability for model 1. These results highlight the importance of external validation. In previous prediction studies of T1 stage ADC with MPP, only He et al. (18) performed external validation, while most others were single-center studies that lacked external validation. There are limitations to single-center studies because of over-fitting in the prediction model. Over-fitting was observed in model 1 after external validation, resulting in poor reliability of study results and limited clinical applicability. Multicenter data must be used for constructing and testing radiomics models for better clinical application.

There were several limitations to this study. First, retrospective studies have inherent weaknesses and potential bias. In future prospective works, we will strengthen the research of ratio of GGO and solid type. Second, the study included a small number of patients with MPP in ADC. Although this study included three hospitals, our sample size was small. This may be because we included patients with peripheral lung ADC with size ≤ 2 cm, and the incidence of MPP in ADC increases with tumor size. A larger sample size is required in future studies. Third, the incidence of MPP in ADC in our study is different from previous studies. As pathologists have different expertise, potential subjective deviation may be present when semi-quantitative analysis is used to record the ratio of each pathological subtype, which may affect the study results.

In summary, the constructed two models based on nodule type stratification has potential to predict MPP in lung ADC of size ≤ 2 cm. We found that GGO nodule type in the without MPP dataset will affect the performance of the prediction model. Thus, the pure solid nodules (model 2) prediction had moderate stable generalizability. This model may contribute to an auxiliary method for preoperative prediction of MPP peripheral lung ADC of size ≤ 2 cm with proper treatment planning.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving human participants were reviewed and approved by the First Affiliated Hospital of Zhejiang University School of Medicine; the First Affiliated Hospital of Wannan Medical College; Zhejiang Cancer Hospital. The ethics committee waived the requirement of written informed consent for participation.

ML and ZF: Study design, manuscript writing and data analysis. ML, YR, and FS: labeled the image. LZ and MW: CT raw data collection. All authors contributed to the article and approved the submitted version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.