Radiomics in the characterization of lipid-poor adrenal adenomas at unenhanced CT: time to look beyond usual density metrics

An adrenal incidentaloma is defined as an asymptomatic adrenal mass discovered on imaging that was not performed to investigate a suspected adrenal disease [1]. In most cases, adrenal incidentalomas represent benign non-functioning adenomas, but they may also correspond to different conditions requiring full clinical attention and therapeutic intervention (e.g., adrenocortical carcinoma, pheochromocytoma, hormone-producing adenoma or metastasis). As a consequence of the burgeoning use of advanced diagnostic imaging in daily medical practice, in the last decades, we have observed a constantly increasing incidence rate of incidentally discovered adrenal nodules. Indeed, adrenal incidentalomas are common, estimated to occur in approximately 3 to 7% of adults [2, 3].

Incidental adrenal masses represent diagnostic challenges for both radiologists and referring clinicians, particularly when the initial imaging features are equivocal or non-diagnostic. The main challenge is correctly identifying the infrequent unexpected malignant lesions (or hyperfunctioning adenomas), while sparing the vast majority of patients with clinically insignificant disease from unnecessary further examinations.

Diagnostic imaging is crucial in the classification of adrenal masses, since the precise etiology can be determined on both computed tomography (CT) and magnetic resonance imaging (MRI) for several entities without the need for further tests [1, 4].

In particular, CT could aid the diagnosis of adrenal adenomas in two ways, namely density measurement and contrast washout. A density lesser than 10 HU on non-enhanced CT (NECT) is almost always diagnostic of a lipid-rich adenoma, regardless of size [2]. By contrast, if there are no benign diagnostic imaging features (for instance, macroscopic fat, adrenal density <10 HU), a dedicated adrenal CT protocol including a 15-min delayed acquisition after contrast media administration is advisable, in order to assess the absolute—or relative—percentage washout. However, pheochromocytomas and adrenal metastases from hypervascular primary extra-adrenal malignancies could sometimes exhibit a washout pattern similar to that of adrenal adenomas [5‐8].

Other imaging modalities may be useful to clarify the nature of the nodule, in particular MRI, in which a signal loss between in- and opposed-phase images at chemical-shift imaging is diagnostic of adenoma, or positron emission tomography (PET)-CT, in which most adenomas show FDG uptake less than 3.1 [9].

However, the need for additional tests puts patients at risk of anxiety and unnecessary harm from diagnostic procedures; additionally, the costs incurred can be significant.

Radiomics refers to a rapidly emerging discipline based on the extraction of mineable data from medical imaging. It has been used in oncology to support diagnosis, prognostication, and clinical decision-making, with the goal of delivering precision medicine [10‐13].

In recent research, O’Shea et al [14] and Cao and Xu [15] demonstrated that early-stage metastases may be differentiated from lipid-poor adenomas using contrast-enhanced CT and NECT radiomics feature–based models with high performances. In other research, radiomics was used to distinguish lipid-poor adenomas from paragangliomas, phrochromocytomas, or carcinomas [16, 17].

To discriminate lipid-poor adenomas from other adrenal lesions, Zhang et al [18] recently developed three prediction models using conventional, radiomics, and combined feature nomograms. However, there was no significant difference in performance between the radiomic and traditional models.

In this study, we retrospectively assessed a dataset of adrenal masses with pathological confirmation that had been classified at NECT as indeterminate and that had not been distinguished by standard clinical demographic or radiological characteristics.

We hypothesized that indeterminate adrenal lesions of benign and malignant nature may exhibit different values of key radiomics features, and we developed a radiomic signature for the classification of benign lipid-poor adenomas, which may potentially help clinicians limit the number of unnecessary investigations in clinical practice.

This retrospective study was conducted according to the Declaration of Helsinki; local Ethics Committee approval for data collection was obtained (Ethics Committee of Area Vasta Emilia Centrale (AVEC); protocol code: 146/2022/Oss/AOUFe, approved on 17/02/2022). All investigations were performed by routine clinical practice and retrospectively retrieved.

Physicians’ desire for diagnostic certainty and, on the other hand, discomfort with diagnostic uncertainty when faced with an unexpected or unexplained imaging finding can lead to an increase in test ordering. As a result, further imaging and clinical evaluation are often performed when an adrenal incidentaloma is discovered and when imaging findings are equivocal or inconclusive [2, 21, 22].

In the present study, a radiomic signature composed of first- and higher-order features, namely quadratic mean, strength, maximum 3D diameter, volume density, and area density, showed a very good average performance with AUC = 0.94 (0.91–0.96) among the four final classifiers to discriminate adenomas from other adrenal lesions at NECT.

The performances of our machine learning models did not differ significantly, with logistic regression showing the best results with an AUC of 0.96 (Fig. 5a). The true positive rate for adenomas, according to the model, was 79%, whereas it was higher (93.5%) for non-adenomas, as shown in Fig. 6a. The decision tree performed better for both classes, with a true positive rate of 84.2% and 90.3% (Fig. 6d).

Our signature’s performance is comparable to that of the three models developed by Zhang et al [18] for differentiating lipid-poor adenomas using conventional, radiomic, and integrated conventional-radiomic CT features. These models had an AUC of 0.94, 0.93, and 0.96, respectively. However, in their cohort, conventional parameters such as gender, age, mean HU, and tumor diameter were strong predictors of the outcome at both univariable and multivariable logistic regression. As a result, their radiomic signature did not significantly improve the performance of the conventional model, i.e., employing standard radiological features and demographic data, thereby diminishing the scientific impact of their results.

Conversely, the net benefit of using our radiomic signature in comparison to standard parameters is clearly visible from the decision curves shown in supplementary figure 2. In our opinion, these differences are easily addressed by the different cohorts of patients used to train and test the models in the two studies, which may have led to some selection biases while considering broader inclusion criteria for eligible lesions in the aforementioned work. As shown in Table 2, indeed, no standard parameters in our cohort were statistically significant predictors of the outcome using multivariable regression.

The composition of the radiomic signature reveals additional distinctions. Indeed, our signature is composed of quadratic mean, strength, maximum 3D diameter, volume density, and area density. The quadratic mean is a first-order feature derived from a histogram that has a fair correlation to the HU median (or mean) value, which is found in the work of Zhang et al [18], Cao and Xu [15], and O’Shea et al [14] for the differentiation of lipid-poor adenomas from other histotypes.

Strength is a more complex second-order feature that is related to the texture of the image; in particular, it can be correlated to the concepts of coarseness, as specifically described by Amadasun and King [23]. In this context, a high strength means that the patterns that compose the texture of the tumor appear larger with broader areas of uniform pixel intensities whereas a low strength would correspond to a finer texture leading to higher variations in local pixel intensities. To our knowledge, this predictor has not been investigated in any other published study.

Maximum 3D diameter is a parameter already employed in clinical practice and reported in previous studies cited above. In the end, area and volume density are related to the shape and extent of the tumors and may provide additional information on their morphological appearance. In fact, adenomas present more frequently as well-demarcated round or oval lesions [9]. These two parameters likely reflect and quantify these visual characteristics of the tumor that were not quantified in previous studies or were only partially considered when the greatest or shortest diameters and tumor volume were used [15]. Our findings suggest that additional metrics, beyond the mere measurement of the mean density, should be considered for inclusion in routine radiological evaluation of adrenal lesions to reduce the number of incidentalomas regarded as indeterminate at NECT examination, thus avoiding unnecessary clinical workup and follow-up examinations.

At NECT, adenomas present more frequently with low attenuation (less than 10 HU) due to a microscopic fat component. This cut-off is highly specific (sensitivity 71%, specificity 98%) [4, 6], widely accepted in the scientific literature, and routinely employed in radiological practice [24, 25]. However, NECT alone is not always diagnostic, since 15–30% of adenomas are lipid-poor, namely containing insufficient intracytoplasmic lipid to conform to the non-contrast features previously described, thereby demonstrating higher attenuation values [26]. Previous works have shown that decreasing the HU threshold for the identification of adenomas could improve the specificity but reduce the sensitivity, whereas increasing such a threshold could result in improved sensitivity but reduced specificity [27, 28].

In a study by Yi and colleagues [17], aiming to differentiate histology-confirmed lipid-poor adenomas from pheochromocytomas, the authors built two radiomic nomograms using NECT and contrast-enhanced CT data, respectively, and concluded that the additional contrast-enhanced adrenal CT may not be necessary. Indeed, the drawbacks of a second scan can include additional costs, radiation risks, and potential harms associated with contrast media administration, including allergy and potential renal injury. A dedicated adrenal CT protocol including a 15-min delayed acquisition and considering a 60% threshold for contrast washout has been shown to properly classify 96% of adrenal masses, with 98% sensitivity and 92% specificity for discriminating adenomas from non-adenomas [29]. However, it should be noted that the additional role of dedicated CT protocols in characterizing incidental adrenal masses based on washout calculation has a low sensitivity and specificity in the literature, particularly in the case of suspected pheochromocytomas or metastases from hypervascular tumors which frequently demonstrate rapid contrast washout [22]. Hypervascular metastases from renal cell carcinoma and hepatocellular carcinoma are examples that may include intracellular lipids and have washout values similar to adenomas [30].

Furthermore, the patient is required to return for dedicated adrenal imaging if the initial NECT, in which the lesion had been incidentally detected, was inconclusive; this will obviously lengthen the diagnostic process and cause psychological distress to the patient.

There are several limitations to this study that should be considered. One major limitation is the small sample size of the final cohort. This was inevitable because our efforts to find patients who had adrenal nodules that were indeterminate at NECT, with the necessity of histological confirmation, resulted in a relatively small number of lesions meeting the inclusion criteria. Another limitation is the retrospective nature of the image data acquisition: in this observational study, the type of scanner used for each patient was not controlled. When considering the robustness of radiomic applications in the clinical setting, the potential impact of variation in CT data acquired from different scanners should not be understated. However, our radiomic signature is based on 1 histogram-based feature, 1 second-order feature, and 3 shape features that have been shown to be robust in previous radiomics studies [31, 32].

Including additional imaging indicators for the identification of lipid-poor adenomas can increase the accuracy of NECT and reduce the need for additional imaging and clinical workup, according to this and other recent studies focusing on radiomics that have distinct points of contact with current clinical practice.

Our radiomic signature based on 1 histogram-based feature, 1 second-order feature, and 3 shape features could be considered for integration in routine radiological assessment of adrenal lesions, beyond the mere measurement of median density. This may serve as a method of enhancing the diagnostic power of NECT in order to substantially limit the number of adrenal incidentalomas initially regarded as indeterminate.