The present study showed that the AIAS has good performance in the qualitative diagnosis of pulmonary nodules, which may be used as an auxiliary diagnostic tool for clinicians to differentiate benign and malignant pulmonary nodules.
Previous studies have shown [13] that segmental resection of lung tissue is mostly used in AIS and MIA, while lobectomy is often used in IAC. In recent years, AIAS has developed rapidly, which has been gradually applied to various fields with deep learning and convolutional neural networks [14-16]. In this study, the clinical pathological diagnosis of patients and AIAS data related to pulmonary nodules were tested for the consistency by Kappa value, which is one of the main innovations of this study. Based on the big data provided by several well-known tertiary hospitals in China, the AIAS is trained with the dataset of pulmonary nodules with clear pathological results as the gold standard. As a result, the probability of malignancy of pulmonary nodules can be estimated. This study indicated that AIAS might have a good coincidence rate on qualitative diagnosis of pulmonary nodules with the pathological, and the Kappa value was 0.809, which was consistent with the results reported by Bejnordi BE et al [17-19]. The ability of AIAS not only to accurately detect and localize pulmonary nodules, but also to accurately predict the probability of malignancy of pulmonary nodules, demonstrates the importance of artificial intelligence in assisting clinicians in diagnosing early-stage lung cancer.
When malignant pulmonary nodules spread and invade the surrounding lung tissue, they are more likely to have rough edges [20], whereas benign nodules usually have smooth surfaces and are clearly demarcated from the surrounding lung tissue. In the present study, there were statistically significant differences between the pathological findings of benign and malignant pulmonary nodules in terms of age, volume and malignant signs on the surface of the nodules, which is consistent with previous studies [21]. Several other studies have reported that the diameter and density of nodules also contribute to risk stratification of pulmonary nodules [19], which is not entirely consistent with our study, and the relevant reasons are considered as follows: (1) the limited source of nodules enrolled in this study and the small sample size may have led to biased results; (2) the enrolled patients were grouped according to the latest version of the Pulmonary Oncology Guidelines. The progression from MIA to IAC is a dynamic development process, with the proliferation and filling of tumor cells, the size of aggregated nodules becomes larger and larger, and during the growth process malignant signs such as burrs, lobulation, and nodular thickening appear on the surface of the lung nodules and accumulate towards the lesion area, characterized by vascular aggregation, fibrosis of the subpleural lesion, and scar contracture is drawn to the visceral pleura to produce pleural depression [22]. The present study showed a 7.983-fold increase in the invasiveness of lung adenocarcinoma when any signs such as irregular lobulation, pleural depressions, burrs of variable length, and vascular aggregation were present on the surface of the lung nodules, which is consistent with the above results.
It was also reported that IAC had more fibroblasts than MIA, which could indicate that the pleural indentation sign was closely related to the degree of infiltration of the lesion [23]. In the present study, the area under the curve for malignant signs was 0.786. When malignant signs were present on the surface of lung nodules, the sensitivity of AI was 71.4% and the specificity was 85.7% in diagnosing the degree of lung adenocarcinoma infiltration. It can be concluded that the specificity of AI is slightly higher than the sensitivity in detecting malignant signs on the surface of lung nodules, which may be due to the more invasive degree of MIA or the weaker defense of normal lung tissue, making the surface of lung nodules irregular.
It has been reported that lung nodule volume based on CT features is an independent risk factor for lung nodule infiltration and volume is an important predictor of tumor growth characteristics and degree of infiltration [24], which is consistent with the present study. The results of this study showed that the AUC of nodal volume was 0.891. When the nodal volume was 748.95 mm3, the sensitivity and specificity of AI in diagnosing the degree of lung adenocarcinoma infiltration was 88.6% and 83.3%. Based on these results, it indicates that AI has high sensitivity and specificity in measuring the volume of lung nodules for diagnosing the extent of lung adenocarcinoma invasion. Therefore, the three-dimensional volume of lung nodules by AI has high diagnostic value for predicting the degree of tumor invasion.
The study has some limitations. First, this study retrospectively analyzed pulmonary nodules with definite pathological findings, but the limited number of cases enrolled in this study and the small sample size may have an impact on the findings. In the future, it is necessary to expand the sample size to include more factors to further explore the efficacy of AIAS for the diagnosis of pulmonary nodules. Second, pulmonary nodules without pathological findings were not included in this study, and there may be some bias in case selection.