Table 1 Summary of studies looking at the application of ML techniques in AI management.

Yi et al.¹⁰

To differentiate between subclinical pheochromocytoma (and lipid-poor adenoma in cases of AI using texture features of unenhanced CT scans

80 patients with lipid-poor adenoma and 29 patients with subclinical s pheochromocytoma

Logistic regression (accuracy of 94%) and number of positive features by comparison to cut-off value (accuracy of 85%)

ML-based quantitative texture analysis on unenhanced CT scans appears to offer a dependable quantitative approach for distinguishing between pheochromocytoma and lipid-poor adenoma in cases of AI

Discrepancy in sample sizes between the two groups. Lack of division of data into training and test datasets. Results for one ML classification method only

Yi et al.¹⁴

To differentiate between subclinical pheochromocytoma and lipid-poor adrenal adenoma in AI using texture and other parameters of CT images

181 patients with lipid-poor adenoma and 84 patients with subclinical pheochromocytoma

Logistic regression using contrast-enhanced CT (AUC of 0.967), and using pre-enhanced CT (AUC of 0.958)

ML approach for pre-enhanced and enhanced CT images distinguished subclinical pheochromocytoma from lipid-poor adenoma. In particular, a good result for CT without contrast allows to avoid the additional radiation and risk associated with enhanced CT

Discrepancy in sample sizes between the two groups. Results for one ML classification method only

Elmohr et al.¹³

To distinguish large adrenal adenomas and carcinomas using texture features of precontrast and venous CT images and tumour attenuation values

25 patients with adrenocortical adenoma and 29 patients with adrenocortical carcinoma

Logistic regression (accuracy of 82%, texture features and attenuation) and Boruta random forest (accuracy of 76%, texture features only)

CT texture analysis of large adrenal tumours and carcinomas is likely to improve CT evaluation of AI

Highly selective nature of the included adrenal tumours

Delayed-phase CT images were not included. Results for one ML classification method only

Liu et al.⁹

To differentiate subclinical pheochromocytoma from lipid-poor adenoma in patients with AI using parameters of pre-enhanced and enhanced CT images analysed by radiologists

183 patients with lipid-poor adenoma and 86 patients with subclinical pheochromocytoma

Logistic regression model (best accuracy of 86%), SVM and Random Forest (lower accuracy than LR, no exact figures were given)

The promising application of CT-based ML models and scoring systems for predicting the histology of AI was demonstrated

Lack of arterial phase and multi-phase scans of CT. Results for one ML classification method only

Maggio et al.¹¹

To differentiate between cortisol secreting and non-secreting AI using texture features of CT scans in non-contrast phase

40 patients with functioning and 32 with non-functioning adrenal masses

Logistic regression (sensitivity of 93.75% and a specificity of 100%)

CT texture analysis shows potential as a valuable tool in defining the diagnosis of AIs

Large number of features incorporated into the predictive model. Results for one ML classification method only

Yang et al.¹⁵

To distinguish between aldosterone-producing adenoma from non-functioning adrenal adenoma using contrast-enhanced CT image features combined with clinical features

68 patients with aldosterone-producing adenoma 60 patients with non-functioning adrenal adenoma

Logistic regression using CT image features (accuracy of 73%) and logistic regression combining CT and clinical features (accuracy of 96%)

Contrast-enhanced CT -based radiomics and clinical radiomics ML model exhibited good diagnostic efficacy in differentiating aldosterone-producing adenoma from non-functioning adrenal adenoma

Only patients with contrast-enhanced CT imaging data were included. Highly selective nature of the included tumours. Results for one ML classification method only

Piskin et al.²¹

To differentiate between nonfunctioning and autonomous cortisol-secreting AI using texture features of unenhanced MRI images

100 patients with adrenal lesions

Logistic regression, best results using MRI image features (AUC of 0.758)

Non-functioning AI and autonomous cortisol-secreting AI can be distinguished with high accuracy on unenhanced MRI

Radiomics analysis and the model built using ML algorithms appear to be superior to radiological assessment method

Results for one ML classification method only

Piskin et al.²²

To differentiate between non-functional adrenal incidentaloma and adrenal Cushing’s syndrome in cases of AI using texture features of MRI

50 patients with AI

Logistic regression (best model AUC 0.994)

The developed MRI-based radiomic scores can yield high area under curves for prediction of adrenal Cushing’s syndrome

The assessment of interobserver reproducibility in feature extraction was not feasible as only one radiologist assessed the images

Results for one ML classification method only

Feliciani et al.¹⁶

To differentiate between pathologically proven adenomas and other adrenal histotypes using texture features of unenhanced CT images

48 patients with 50 adrenal lesions

Four classifiers were used: logistic regression (AUC of 0.96), linear discriminant (AUC of 0.95), linear SVM (AUC of 0.94), decision tree (AUC of 0.91)

The research constructed a radiomic signature based on unenhanced CT scans to categorize lipid-poor adenomas

Lack of control over CT scanner types due to the retrospective nature of the study