CT radiomics including lesion-surrounding regions for distinguishing pulmonary cryptococcosis from lung cancer

Pulmonary cryptococcosis (PC) is a fungal lung disease, and nodule/mass-type PC can exhibit imaging findings similar to lung cancer (LC). This study aimed to develop and externally test a new radiomics model based on the lesion and lesion-surrounding regions for distinguishing PC from LC.

In this retrospective study, patients with either PC or LC who underwent non-enhanced CT at four hospitals were included. A considerable number of radiomics features were extracted from the lesions and their surrounding regions (0–1 mm, 0–2 mm, and 0–3 mm). Ten methods were used to calculate Rad-score, followed by a 10-fold cross-validation. A combined model was developed by integrating Rad-score and clinical factors. The models were subsequently tested using external test sets and compared in terms of the area under the curve (AUC).

A total of 391 patients, including 159 with PC and 232 with LC, were enrolled in this study. These patients were divided into three groups: training set, external test set 1, and external test set 2. In terms of Rad-score, linear SVC demonstrated the highest AUC of 0.901 in cross-validation. For the combined model, logistic regression showed the best predictive performance with an AUC of 0.946 in the cross-validation. A lesion-surrounding feature (0–3 mm) played the biggest role in both Rad-score and the combined model, accounting for the highest relative weight. In the external tests, the combined model exhibited higher AUCs (0.872 in the external test set 1, and 0.952 in the external test set 2) compared to Rad-score and the clinical model.

The combined model can aid in differentiating PC from LC, with the lesion-surrounding radiomics playing the most significant role.