Machine learning analysis of adrenal lesions: preliminary study evaluating texture analysis in the differentiation of adrenal lesions

Abstract

Purpose: This study aimed to determine the accuracy of texture analysis in differentiating adrenal lesions on unenhanced computed tomography (CT) images.

Methods: In this single-center retrospective study, 166 adrenal lesions in 140 patients (64 women, 76 men; mean age 56.58 ± 13.65 years) were evaluated between January 2015 and December 2019. The lesions consisted of 54 lipid-rich adrenal adenomas, 37 lipid-poor adrenal adenomas (LPAs), 56 adrenal metastases (ADM), and 19 adrenal pheochromocytomas (APhs). Each adrenal lesion was segmented by manually contouring the borders of the lesion on unenhanced CT images. A texture analysis of the CT images was performed using Local Image Feature Extraction software. First-order and second-order texture parameters were assessed, and 45 features were extracted from each lesion. One-Way analysis of variance with Bonferroni correction and the Mann-Whitney U test was performed to determine the relationships between the texture features and adrenal lesions. Receiver operating characteristic curves were performed for lesion discrimination based on the texture features. Logistic regression analysis was used to generate logistic models, including only the texture parameters with a high-class separation capacity (i.e., P < 0.050). SPSS software was used for all statistical analyses.

Results: First-order and second-order texture parameters were identified as significant factors capable of differentiating among the four lesion types (P < 0.050). The logistic models were evaluated to ascertain the relationships between LPA and ADM, LPA and APh, and ADM and APh. The sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy of the first model (LPA vs. ADM) were 85.7%, 70.3%, 81.3%, 76.4%, and 79.5%, respectively. The sensitivity, specificity, PPV, NPV, and accuracy of the second model (LPA vs. APh) were all 100%. The sensitivity, specificity, PPV, NPV, and accuracy of the third model (ADM vs. APh) were 87.5%, 82%, 36.8%, 98.2%, and 82.7%, respectively.

Conclusion: Texture features may help in the characterization of adrenal lesions on unenhanced CT images.

Keywords: Adrenal adenoma; adrenal glands; adrenal mass; computed tomography; texture analysis.

Figure 1

Flowchart of patients with adrenal lesions. CT, computed tomography; MRI, magnetic resonance imaging; PET-CT, positron emission tomography-computed tomographöy.

Figure 2

(a) Lipid-rich adenoma in a 49-year-old woman. A well-circumscribed and low-density lesion was observed in the right adrenal gland in axial computed tomography (CT) (white arrows). (b) A 37-year-old woman with lipid-poor adenoma diagnosed histopathologically. Axial unenhanced CT showed a lipid-poor right adrenal mass (white arrows). The mass with a density of 22–56 Hounsfield unit (HU) had a heterogeneous internal structure. (c) Adrenal metastases were diagnosed through surgery in a 56-year-old man with lung cancer. Axial unenhanced CT showed a large, well-defined, homogenous left adrenal mass with a density of 54 HU (white arrows). (d) Surgically resected pheochromocytoma in a 32-year-old man. Axial unenhanced CT showed a well-defined, homogenous, spherical left adrenal mass with a density of 46 HU (white arrows).

Figure 3

Segmentation examples of reader 1 (a) and reader 2 (b). (a) Texture analyses of the lipid-rich adrenal adenoma using local image feature extraction (LIFEx) software [lesion tagged with a yellow region of interest (ROI)]. (b) Texture analyses of the lipid-poor adrenal adenoma using LIFEx software (lesion tagged with a green ROI). (c) Texture analyses of the adrenal metastasis using LIFEx software (lesion tagged with a red ROI). (d) Texture analyses of the adrenal pheochromocytoma using LIFEx software (lesion tagged with a pink ROI).