With the development of molecular pathology in recent decades, PSC, an uncommon but highly malignant tumor, has been further recognized. The latest 2021 WHO classification of lung neoplasms categorized PSC into three pathological subtypes: pleomorphic carcinoma (giant cells and/or spindle cells > 10%), carcinosarcoma, and pulmonary blastoma [17]. Pleomorphic carcinomas have striking immunohistochemistry and molecular pathology similarities with conventional NSCLC, and their mesenchymal components express epithelial immunohistochemical markers, supporting the notion that PSC is a transformed or dedifferentiated variant of conventional NSCLC [2, 18, 19]. The sarcomatous components of carcinosarcoma are real sarcomas, and immunohistochemistry analysis of the sarcomatous components does not reveal the expression of epithelial markers. A dual-stage pulmonary blastoma is a neoplasm composed of embryonic epithelial elements and primitive mesenchymal stroma. Although carcinosarcoma and pulmonary blastoma have biphasic histological components, recent studies have found that both components have the same genetic mutations, strongly suggesting that the tumors are of monoclonal origin and are differentiated from totipotent tumor stem cells that underwent neoplastic transformation with epithelial and sarcomatous cell differentiation [2, 20–22]. Regardless of tumor composition, PSC is often found in an advanced stage with high mitotic activity, vascular infiltration, hemorrhage, necrotic areas, and pleural invasion [2].
In the present study, PSC patients were more likely to be male smokers, with a male-to-female ratio of 5.2:1. The average age at diagnosis was 63 years, and the average tumor size was 5 cm, which was similar to the literature reports [4, 10, 16, 23]. Our study showed that the median survival time and the one-year and three-year OS rates of PSC were shorter than those of other NSCLC with comparable sizes. This indicates that PSC is a special subtype of NSCLC that is associated with worse outcomes.
In this study, the CT findings showed that calcification and vacuoles/cavities were rare in PSC, consistent with a previous study [23, 24]. However, Ung and colleagues [10] found that 20% of their cases had irregular cavities, which may be related to the small number of cases in our study. PSCs are prone to pleural invasion, and Kim and colleagues [8] showed that pleural invasion occurred in 7 out of 10 cases (70%). In the study of Fujisaki and colleagues [16], the pleural invasion rate was 43% (19/44). In our study, this rate was 45.2% (14/31), while other NSCLC with comparable sizes also had a high proportion of pleural invasion (36/56, 64.3%), suggesting that this CT finding is not specific—pleural invasion is not surprising in malignant tumors when the tumor is sufficiently large. Therefore, conventional CT findings are difficult to diagnose PSC preoperatively.
This study found that PSC had a larger LAA ratio compared to other NSCLC, with an optimal cutoff value of 16.6%. The pathological components of the low-attenuation areas of PSC were mucinous degeneration, necrosis, and hemorrhage [8], suggesting that the rapid proliferation of tumor cells exceeded the blood supply [24]. Such areas were found to be an independent predictor of PSC prognosis: patients with a larger LAA (> 25%) had a shorter overall/disease-free survival than that of smaller LAA (< 25%) [16].
Besides, this study found that CT radiomics could distinguish PSC from other NSCLC, among which PS alone, PS + AP, and PS + VP had higher diagnostic efficacy, and PS had the highest diagnostic efficacy. Sha and colleagues [1] applied CT radiomics features to discriminate mediastinal lymph node metastases of NSCLC, and the results were similar to ours: both single-phase and combined-phase models achieved good diagnostic performance, but PS had the highest diagnostic performance. A recent study on the differential diagnosis of benign and malignant parotid gland tumors also showed that plain scanning can achieve high differential diagnosis efficiency [25]. The diagnostic performance of the single enhanced sequence in this study was relatively low, which may be because PS reflects the original heterogeneity of tumors, while enhancement of supplying vessels and inhomogeneous tumor enhancement reduces the differences between the two kinds of tumors.
In addition, this study showed that LR, SVM and RF were better algorithms for differentiating between PSC and other NSCLC. Our study is similar to the report from Han and colleagues [26], in which RF, LDA and SVM were better algorithms for identifying adenocarcinoma and squamous cell carcinoma. Qian and colleagues [27] indicated that LASSO + SVM was the best combination for distinguishing glioblastoma from metastasis in the brain. Bi and colleagues [28] recently found that the LASSO + LR algorithm yielded the highest performance in differentiating pancreatobiliary and intestinal-type periampullary carcinomas.
However, this study also has several limitations. Firstly, the number of PSC cases in this study was small, even though the cases dated back 10 years. Due to the rarity of PSC, only multicenter studies can guarantee a larger number of cases. Secondly, this study has not explored whether CT radiomics can predict postoperative recurrence of PSC and evaluate its treatment response.
In conclusion, this study explored the feasibility of preoperative differentiation of PSC from other NSCLC based on CT findings and radiomics. We found that PSC had a larger LAA ratio compared to other NSCLC, with an optimal cutoff value of 16.6%. In CT radiomics, the combination of PS data and LR algorithm yielded the highest diagnostic efficacy. These findings would be helpful for the accurate preoperative diagnosis and clinical treatment of PSC.