Cloud Hosting

Aim

An adrenal incidentaloma (AI) is an asymptomatic adrenal mass that is recognized incidentally during imaging examinations and is not associated with suspected adrenal pathology1,2. Incidental discovery of adrenal masses has increased recently due to wider application and technical improvement of abdominal imaging procedures, with a prevalence of approximately 0.26.9% in radiological studies1,3,4,5. A comprehensive hormonal evaluation of newly diagnosed adrenal masses at their initial presentation was recommended by the European Society of Endocrinology in 20166.

Patients should be referred for adrenalectomy with clinically significant hormone excess, radiological findings suspicious for malignancy, signs of local invasion, and when the tumour is greater than 5cm6. Underlying comorbidities, advanced age, and Hispanic ethnicity were associated with more frequent postoperative complications. Therefore, the coexistence of heart failure or respiratory failure should always be considered before qualifying for surgical treatment of adrenal tumours7.

The primary objective of this study was to compare several machine learning (ML) techniques in a qualification for adrenalectomy and choose the most accurate algorithm as a valuable adjunct tool for doctors to simplify making therapeutic decisions by using the most innovative and modern methods. To the best of our knowledge, this study is the firstattempt to apply ML techniques to qualify for the surgical treatment of AI using both the results of diagnostic tests and computed tomography (CT) image features. Preliminary results of this study were presented in a poster session at the European Congress of Endocrinology8.

In the literature, most studies apply computer vision techniques to recognize the type of tumour based on CT images9,10,11,12,13,14,15,16. In one study, the authors evaluated ML-based texture analysis of unenhanced CT images in differentiating pheochromocytoma from lipid-poor adenoma in adrenal incidentaloma10. The textural features were computed using the MaZda software package, and two classification methods were used: multivariable logistic regression (accuracy of 94%) and number of positive features by comparison to cut-off values (accuracy of 85%). The results were encouraging; however, decision classes were unbalanced and the accuracy values were computed on the test set. Therefore, they were biased estimators. In another study, the authors applied a multivariable logistic regression model with 11 selected textural features computed using MaZda software11. The cut-off point obtained using the eceiver operating characteristic (ROC) curve applied to the expression obtained from logistic regression resulted in a sensitivity of 93% and 100% specificity. Again, these results were obtained using the same set used to train the model. In another study performed by Li et al., ML models were used to differentiate pheochromocytoma from lipid-poor adenoma based on the radiologists description of unenhanced and enhanced CT images9. The authors used three classifiers: multivariate logistic regression, SVM and random forest. As a result, two separate models based on multivariable logistic regression were proposed, each using three CT features: M1 with preenhanced CT value, shape, and necrosis/cystic changes (accuracy of 86%) and M2 using only preenhanced CT features: CT value, shape, and homogeneity (accuracy of 83%). Elmohr et al. used the ML algorithm to differentiate large adrenal adenomas from carcinomas on contrast-enhanced computed tomography, and its diagnostic accuracy for carcinomas was higher than that of radiologists13. Other studies have evaluated the accuracy of ML-based texture analysis of unenhanced CT images in differentiating lipid-poor adenoma from pheochromocytoma, with performance accuracy ranging from 85 to 89%10,14.

The literature also includes papers applying ML techniques to magnetic resonance imaging (MRI) data. An example of such work is a study where the authors utilized logistic regression with the least absolute shrinkage and selection operator (LASSO) to select MRI image features and distinguish between non-functional AI and adrenal Cushings syndrome17.

In studies involving a large number of features (e.g.: software packages such as MaZdA can calculate several hundred texture parameters for images), dimensionality reduction is required. Techniques commonly used (or combinations of them) are: LASSO with regression18,19,20,21, elimination of correlated features9,21 or those with low intraclass correlation (ICC)18, training of classifiers for subsets of features and selection of subsets with the highest classifier accuracy9, elimination of features with p-values above the accepted error rate for coefficients in regression models, use of feature discrimination power calculated using the ROC curve for each feature separately10.

Artificial neural networks (ANN) are flexible and powerful ML techniques that have evolved from the idea of simulating the human brain, however their successful application usually requires datasets much larger that other classification methods17,18,19.

To improve the quality of patient care, recent studies have been conducted in several different sectors using modern techniques. There are two types of ML-based models: current-condition identification and forward prediction20. In Table 1, we have summarized studies concerning the utilization of ML techniques in AI management.

Read more from the original source:
Patient classification and attribute assessment based on machine learning techniques in the qualification process for ... - Nature.com