The application of big-data to the study of texture analysis enables the discovery of novel predictive factors in oncology [20]. Radiomics leverages radiological images acquired in routine diagnostic practice, using the data to perform complex mathematical calculations. The images analyzed can be composed of pixels, the conventional unit of measurement for the surface of a digital image, or voxels, the three-dimensional counterpart of pixels [30]. Each radiological image can be seen as a set of pixels that combine precise spatial coordinates and the intensity of the grayscale level. The concept of texture in a digital image, therefore, can be understood as the distribution and ratio of intensity levels in the pixels of a given ROI or VOI: texture is a characteristic of a region of interest. To describe the relationships between the intensity values of the grayscale in the images, statistical models are generally used, employing dedicated tools and software [8]. The radiomic data thus obtained are classified into four main categories: data related to the shape of the object under examination, first- and second-order statistical data, and higher-order statistical data [8, 31].

The present study aims to determine whether texture parameters derived from CT scans of the target lesion with larger dimensions in patients with advanced ACC before EDP-M treatment are independently associated with the response to chemotherapy (response defined by two radiologists based on the simultaneous agreement of RECIST, Choi, and volumetric criteria). Two parameters could discriminate between responders and non-responders patients: they have been identified after the extraction of radiomic parameters of first-and second-order for the largest primary/metastatic mass of each patient and comparing these parameters based on the binary assignment of patients (assignment defined based on the aforementioned gold standard). This was observed both when considered individually and simultaneously.

To fully understand the clinical significance of these findings, it is necessary to delve into the meaning of the two parameters under examination. Histogram Intensity Kurtosis measures the weight of the tails of the distribution in terms of frequency, defining them as “heavy” or “light” compared to a normal distribution and how much differs from it [29]. Specifically, distributions with low kurtosis tend to have light tails and thus few outliers (interpreted as anomalous values, distant from the other available observations), conversely, higher kurtosis implies greater importance of the tails of the distribution in terms of frequency. In a normal (Gaussian) distribution, kurtosis has a value of 3. Kurtosis excess, with respect to a normal distribution, is the value k – 3. Often, kurtosis is referred to k excess (thus k − 3), hence the k excess of a normal distribution is 0. A distribution with k > 3 (i.e., k excess > 0) is called leptokurtic, while one with k < 3 (k excess < 0) is called platykurtic. The Long Run High Grey-Level Emphasis (LRHGLE) is a parameter derived from the Grey-Level Run-Length Matrix (GLRLM), which is a texture feature that provides spatial information about the distribution of gray values in ROIs or VOIs, which are defined as sets of pixels (or voxels). The GLRLM is calculated for 13 different directions in 3D (or 4 in 2D), and for each of the eleven texture indices derived from it, the 3D value is the average calculated across the 13 directions (in 3D) and the 4 dimensions (in 2D) [29]. The two elements (i, j) of the GLRLM correspond to the number of homogeneous runs of voxels (j) with intensity (i) in an image. A “run” refers to the length of consecutive voxels that share the same gray level intensity along a specific direction. The GLRLM provides higher-order statistical information and expresses how frequently in the image it is possible to observe these sequences of elements with a certain length and intensity [30]. By utilizing the GLRLM, the distribution of runs of gray levels is described, where the length of the run (run-length) measures contiguous gray levels along a specific orientation.

Therefore, by characterizing the pixels (or voxels) of the tumor, the GLRLM can provide information about the regional heterogeneity of the tumor. Specifically, finer textures tend to have shorter runs, while coarser textures will exhibit longer runs. This matrix allows for the estimation of numerous descriptors, and in our specific case, GLRLM_LRHGLE represents the value of the distribution of long runs with high grey-level intensity. The two parameters that discriminates between responders and non-responders had a higher value in the responder’s group. Therefore, patients who will respond to chemotherapy present a higher level of kurtosis of the histogram of the distribution frequency of grayscale values compared to non-responders. At the same time, they exhibit longer runs of high grayscale intensity voxels, meaning a coarser and more intense tissue texture. Although these two parameters only partially describe the characteristics of both the curve of the frequency distribution of grayscale intensities and their regional relationships in space, the underlying trend favors a coarser and more heterogeneous tissue texture in responder patients compared to non-responders (two example cases are shown in Fig. 2). At the best of our knowledge, in literature there are no studies on CT texture analysis and response to chemotherapy of ACC. In oncological imaging, several studies have highlighted the relationship between tissue heterogeneity features and the malignancy of the lesion, the probability of tumor recurrence, the reduced therapeutic response rate, and a poor prognosis [11, 32, 33]. However, contradictory results have also been found in studies on both primary colorectal cancer and metastatic cancer [32], as well as in a study by Yun et al. that analyzed radiomic data derived from contrast-enhanced CT scans in patients with pancreatic cancer. They reported an association between features suggesting more homogeneous textures and a poorer therapeutic response, suggesting a more aggressive tumor behavior [34]. Our study therefore confirms similar results: patients with advanced and metastatic ACC who would be more likely to respond to EDP-M therapy exhibited texture features consistent with greater heterogeneity and coarseness of tissue texture. Furthermore, the coarseness of the texture itself, with high-intensity grayscale voxels, may reflect a faster metabolism and cell growth, which in turn could be the basis for a better response to chemotherapy. To conclude, the correlation between these CT texture features and treatment response could suggest the future use of texture analysis as a prognostic biomarker in metastatic ACC patients.

Contrast enhanced CT scans in venous phase EDP-M response in homogeneous and heterogeneous ACC. Responder patient: panel a) staging scan showing a large left adrenal lesion (arrowhead) with coarse and inhomogeneous texture (with high values of GLRLM_LRHGLE and Kurtosis), large loco-regional nodal metastases (asterisk) and liver metastases (arrows); panel b) restaging scan after EDP-M therapy with a reduction of the adrenal lesion (arrowhead) and of the locoregional metastases (asterisk) and disappearance of the liver metastases. Non-responder patient: panel c) staging scan showing a large left adrenal lesion (arrowhead) with fine and homogeneous texture (with low values of GLRLM_LRHGLE and Kurtosis) and a thrombus in the inferior vena cava (asterisk); panel d) restaging scan after EDP-M therapy with a slight reduction of the adrenal lesion (arrowhead) and the disappearance of the inferior vena cava thrombus but with the comparison of multiple liver metastases (arrows)

We can also speculate that a greater tissue heterogeneity in responders may be secondary to a mitotane effect, as shown by both the higher presence of outliers in the distribution of tissue grayscale levels (due to more prominent areas of cellular necrosis) and the presence of longer runs of high-intensity gray-scale in the texture, or to a coarse presence of more functionally active cell populations that may then respond better to the EDP-M regimen.

The retrospective design, the limited number of paired patients available, and their different clinical histories (in terms of staging, mitotic index, EDP-M treatment after mitotane progression) are the main limitations of our study. Moreover, in some cases, histological confirmation was not available (as in patient number 11, with severe Cushing’s syndrome and virilization that allowed us to diagnose metastatic ACC) or the Weiss score was not calculated if surgery was performed in another center, in case of insufficient material (due to large amount of necrotic tissue), or if specimen obtained with a biopsy was not sufficient to calculate the score. A multicenter study, with a larger number of recruited patients, will be useful to consider if some known prognostic markers (tumor stage, resection, Ki-67) can be combined with radiomic to enhance its predictive role.

Even if it can be perceived as a weakness, we believe that comparing the evaluations of two radiologists provided us with a stronger standard of reference for classifying treatment response. The high Cohen’s Kappa obtained showed that the measurements by the two radiologists were reproducible. This is important because literature reports that in cases of complex tumor shapes, the measurements of the maximum axis, densitometry, and volume of the lesions can vary significantly among readers [35].

It should be underscored that the standard of reference we used required us to dichotomize the patients into responders and non-responders, without considering mixed responses to chemotherapy or the different biological behaviors of ACCs. We cannot exclude that this classification could have brought to biases in the correlation with the textural features. Nevertheless, we believe this was the most effective standard for this type of evaluation, because obtaining a histological specimen from all lesions was not feasible.

In conclusion, the higher value of kurtosis and the presence of long runs of high gray-scale intensity, obtained from the texture analysis of primary or metastatic ACC lesions in routine CT scans, may predict the response to EDP-M therapy in patients with advanced and/or metastatic ACC and could be used as a prognostic biomarker. Actually, peculiar molecular alteration profiles identified in ACC may represent new targetable events, with new or existent drugs [36]. However, if validated by further studies, radiomics and texture analysis could be used as an independent parameter to predict the response to chemotherapy in patients with advanced ACC.