The potential of Radiomics for the differential diagnosis of faat-poor angiomyolipomas and other renal tumors: a Rewiev

Full Text

**1. Introduction**

Kidney cancer ranks 14th in incidence and 16th in mortality among all malignant neoplasms [2]. Advances in modern radiological imaging have led to an increased detection rate of incidental renal masses, including small-sized ones (<4 cm), over recent decades [3]. Small renal masses (SRMs) are contrast-enhancing lesions with a maximum diameter of ≤4 cm, typically corresponding to stage T1a renal cell carcinoma [4]. While most patients with such renal tumors undergo partial resection or radical nephrectomy, up to 20% of small masses may be benign. Among the latter, lipid-poor angiomyolipoma is the most frequently encountered [5]. Differentiating angiomyolipoma from renal cell carcinoma using radiological methods can be challenging, especially when characteristic macroscopic fat components are absent in its structure [6].

The gold standard for verifying a renal tumor is percutaneous biopsy. However, this invasive procedure can lead to complications such as bleeding and phlebitis. Furthermore, performing this procedure entails additional economic costs.

Radiomics is a promising and rapidly evolving research field aimed at extracting quantitative metrics (radiomic features) from medical images. Radiomic features capture tissue and lesion characteristics and can be used alone or combined with clinical and biological data (histological, genomic) to determine the optimal treatment strategy. Thus, it can be used to characterize a lesion and assess the probability of malignancy before deciding on a treatment plan [7]. A primary current objective of texture analysis in renal tumors is increasing the diagnostic accuracy of determining the histological type, especially for small-sized tumors. The results of non-invasive preoperative texture analysis could help select the optimal management strategy for patients with this pathology.

A limitation of radiomics is its high sensitivity to variations in the parameters used to obtain texture characteristics, which is why strict study protocol adherence is of particular importance.

The aim of this systematic review is to systematize the latest data on the methodology and results of radiomic analysis for the differential diagnosis of lipid-poor AML (lpAML) from other renal tumors (including patients with clear cell renal cell carcinoma) based on CT and MRI images.

**Materials and Methods**

**2.1 Study Protocol**

The study design adhered to the PRISMA (Preferred Reporting Items for Systematic reviews and Meta-Analyses) guidelines (https://www.prisma-statement.org/prisma-2020-checklist).

Literature search, study selection, data extraction, and quality assessment were performed and verified by two independent experts. Any disagreements at these stages were resolved through discussion or consultation with the review team.

**2.2 Literature Search and Study Selection Using Inclusion and Exclusion Criteria**

A search for relevant original research articles was conducted in the PubMed/MEDLINE, Embase, and Cochrane Central Register of Controlled Trials (ClinicalTrials.gov) databases. To systematize current information in the field of radiomics, it was decided to select the most recent articles published from January 1, 2021, to 2025. The final search date was October 5, 2025.

Full original articles in English that met the following PICOS criteria were included:

*   **P (population):** Patients (over 18 years old) with verified lipid-poor angiomyolipoma and renal cell carcinoma (including clear cell renal cell carcinoma).
*   **I (interventions):** Diagnostic studies (CT or MRI) on which radiomic (texture) analysis was performed.
*   **C (comparator):** The control group included patients with renal masses of uncertain nature detected on CT/MRI, which were subsequently confirmed as benign by histopathological examination.
*   **O (outcome):** Diagnostic accuracy (sensitivity, specificity, positive predictive value [PPV], negative predictive value [NPV], area under the ROC curve [AUC]) for differentiating lipid-poor AML from malignant renal masses.
*   **S (study design):** Studies published in English in the last 5 years (from January 1, 2021, to 2025) including patients with histologically confirmed renal cell carcinoma (RCC) and lipid-poor AML, aiming not merely at distinguishing benign from malignant tumors based on radiomics but at subtype-specific diagnosis (i.e., providing diagnostic accuracy specifically for differentiating lpAML from other tumors, including ccRCC, rather than lumped together with other benign renal tumors).

**Exclusion Criteria:**

1.  Articles not related to renal tumors, not using radiomics methods, or not including comparative patient groups with lpAML and ccRCC (including articles where radiomics analysis results for lpAML were not differentiated from the total pool of benign tumors or from AML with normal fat content) and articles where the RCC group did not include patients with ccRCC.
2.  Non-original articles (reviews, meta-analyses, and case reports).
3.  Articles where radiomic features were extracted from other diagnostic images (ultrasound images).

An additional analysis of reference lists from included studies to identify publications suitable for this review ('snowballing') was not performed.

**2.3. Study Selection**

Searches in each source (PubMed/MEDLINE, Embase, ClinicalTrials.gov) were conducted for English-language publications released between January 1, 2021, and October 5, 2025.

The search strategy used both MeSH/EMTREE terms and keyword combinations:
*   "Angiomyolipoma"[MeSH Terms] AND "Radiomics"[MeSH Terms];
*   "Angiomyolipoma"[MeSH Terms] AND "machine learning"[MeSH Terms];
*   "Angiomyolipoma"[MeSH Terms] AND Texture analysis;
*   "Radiomics"[MeSH Terms] AND "Small renal masses".

Titles and abstracts of the retrieved articles were independently screened by two experts; those not meeting the inclusion criteria were excluded. Disagreements at each stage were resolved by a third party overseeing the systematic review process, who made the final decision. This resulted in a selection of articles whose full texts were retrieved for detailed assessment. Following this initial stage, two authors independently evaluated the full texts of all remaining articles for final inclusion or exclusion in the study.

**2.4. Data Collection**

Data were independently extracted by two experts after a thorough assessment of the full text of each article.

The following data were extracted from each study:
*   Author and title;
*   Year of publication;
*   Study objectives and reference standard ('gold standard');
*   Study design and number of included tumors (the number of patients is specified where it differed from the number of tumors);
*   Diagnostic modality with specified scan phases (for CT studies) or pulse sequences (for MRI studies);
*   Method of radiomic feature analysis:
    *   Whether machine learning (ML) was used, names of ML types used; for 'handcrafted radiomics' – statistical methods used for feature selection;
    *   Segmentation dimension (2D/3D) and methods (manual/automatic/mixed);
    *   Best results of the radiomic analysis, indicating AUC, sensitivity, and specificity (if available);
    *   Distribution of radiomic features by class (whether assessed); if assessed, which specific classes were mentioned: shape parameters (2D and 3D); first-order features; second-order features: texture features with several subgroups (Gray Level Co-occurrence Matrix, GLCM; Gray Level Run Length Matrix, GLRLM; Gray Level Size Zone Matrix, GLSZM; Neighbouring Gray Tone Difference Matrix, NGTDM; Gray Level Dependence Matrix, GLDM);
    *   Radiomic features selected by the authors as predictive, along with their significance (if available);
    *   Presence/absence of external validation of the study results.

**2.5. Quality Assessment and Risk of Bias**

Modified questions from the Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2) criteria were used, as there is no ready-made quality assessment tool specifically designed for studies using machine learning, which is increasingly implemented in radiomics research.

The risk of bias was assessed independently by two authors using the QUADAS-2 tool (https://www.latitudes-network.org/tool/quadas-2/). Disagreements were resolved by a third party. This tool allows for the assessment of the risk of bias and applicability across four key domains (see Table 2 in the 'Results' section).

**2.6. Assessment of Heterogeneity of Included Studies**

Nine articles were selected for the review. We did not perform a quantitative assessment of heterogeneity sources in the systematic review due to the limited number of studies and insufficiently detailed descriptions within them. We assume that the following factors could potentially cause heterogeneity in these studies, including different CT and MRI protocols, inclusion and exclusion criteria, varying experience of the specialists assessing the images, and different versions of machine learning algorithms, etc. We plan to investigate whether these factors are sources of heterogeneity when a sufficient number of studies with detailed descriptions become available.

**Results**

**3.1–3.3 Search and Selection of Studies**

Initial search queries were run separately in each database/registry using the aforementioned MeSH terms/keywords. Duplicate publications were removed using Mendeley software, first within the bibliography of each source and then across all three sources combined.

The initial search across all sources (2 databases, 1 registry) using the specified terms identified 792 records. After removing 8 duplicates (across all sources) and excluding 106 records deemed irrelevant based on the absence of an author name (in Embase), 678 works remained for title and abstract analysis. Subsequently, 636 records were excluded as unsuitable (reviews, meta-analyses, focusing on other cancer types, not using radiomic analysis). Forty-one publications were included for full-text assessment based on title and abstract screening. At the next stage, 25 articles were excluded for lacking relevant data, and 7 publications were unavailable for data extraction. Thus, 9 articles were included in the final analysis.

Separate screening forms were created for each selection stage. The overall review process flowchart according to the PRISMA statement is presented in Figure 1.

*Figure 1. Flowchart of the systematic literature search (see Appendix).*

**3.4. Data Systematization**

Data from the articles included in the systematic review were extracted after a thorough assessment of the full text and are presented in a table (Table 1).

*Table 1. Systematized data of publications selected for the systematic review (see Appendix).*

The results of each study included in the systematic review demonstrated statistically significant AUC values for the predictive model based on radiomic features, including excellent and outstanding discriminatory ability.

It is important to note that the mention of statistically significant second-order texture features (including subgroups such as GLCM, GLRLM, GLSZM, GLDM, NGTDM) predominated over first-order histogram features: second-order features were highlighted in 6 out of the 9 reviewed articles. The names of the statistically significant parameters selected for each article can be found in Table 1, section 'Best Radiomic Parameters'. It should be emphasized that texture features belonging to the second-order group analyze the spatial relationship and interdependence of pixels with each other, allowing for a quantitative description of tissue texture and heterogeneity.

In one other article (Wei et al., 2025), first-order statistical parameters were extracted from CT images; among them, Perc.25 was the best parameter. However, it should be noted that this publication did not mention the extraction of texture features (second-order and higher) from the images. In the remaining 2 articles (Chai et al., Bang S. et al.), the names of statistically significant radiomic features or their classes were not provided.

In every study included in the systematic review, the authors utilized machine learning methods. The most common machine learning algorithm was logistic regression (LR) - mentioned in 7 out of 9 publications. In the articles by Han et al. and Bang S. et al., five different machine learning algorithms were compared; the best classification performance was demonstrated by a model based on MLP in the first case and based on XGBoost in the second.

Analysis of the data in this review revealed that 6 out of the 9 included articles used CT data to build predictive models, consistent with clinical practice where multiphase abdominal CT is recognized as the 'gold standard' for detecting and differentiating renal tumors due to its superiority over ultrasound and greater availability compared to MRI [8].

In the other 3 publications, radiomic features were extracted from MR images; notably, Jian et al. used IVIM-DWI in their work in addition to standard MR sequences. IVIM-DWI can differentiate between water molecule diffusion and microcirculatory perfusion, thereby providing a more accurate description of functional cellular changes and tissue microstructure compared to DWI sequences.

Authors in 6 out of the 9 articles included in the systematic analysis set the differential diagnosis of lpAML specifically against the clear cell type of renal cell carcinoma as their research goal, since this subtype is associated with the worst prognosis among all RCC subtypes [9].

**3.5 Quality Assessment and Risk of Bias**

The quality of the included studies was assessed using the QUADAS-2 checklist; results are presented in Table 2. Overall, the quality of the included publications was satisfactory. All publications included in the systematic review are typical examples of case-control studies, which characterizes the risk of bias as high in Domain 1 (Patient Selection).

In 5 out of 9 publications, there is no explicit information on whether a consecutive or random sample of patients was included (they report only the inclusion time frame but do not specify if it was consecutive or random), which may indicate selection bias.

In the article by Bang S. et al. [10] (see Table 2), a critically important methodological shortcoming is noted: the radiologist who verified and refined the segmentation (which is the basis for the index test) had access to the initial diagnosis, meaning the diagnosis was not 'blinded' for the index test evaluation - the study had a high risk of bias that could have led to overestimated diagnostic accuracy measures. Furthermore, several articles lack a clear statement on whether the index tests (radiomic analysis) were interpreted without knowledge of the reference standard results (histology) – i.e., a 'blinded' method, which is a potential source of bias.

The main issues leading to reduced study quality were the absence of external validation of results (87.5% of analyzed articles; external validation was described only in the article by Matsumoto et al. [11] – cross-validation was performed in some studies) and the unavailability or lack of transparency of research data (82%).

*Table 2. Quality assessment of publications and risk of bias according to QUADAS-2 (see Appendix).*

**Limitations**

Despite growing interest in radiomics, the literature review at the article selection stage already showed that most current radiomics research focuses on the basic task of binary classification – distinguishing benign from malignant tumors in general, without further stratification into histological subtypes. Existing models often do not account for the histological diversity of tumors, limiting their clinical applicability for more precise, subtype-specific diagnosis and indicating the need for developing more detailed models.

All studies included in the systematic review have a retrospective design.

Furthermore, differentiating the lpAML subtype is a diagnostically challenging task due to the limited number of cases. As emphasized in their systematic review by Dehghani Firouzabadi F. et al. [12], previous CT-based radiomics studies aimed at distinguishing lpAML from clear cell RCC also consistently faced an insufficient number of lpAML cases, including only between 10 and 50 patients.

Thus, an important limitation in several articles was the small sample size for patients with lpAML (the minimum number of lpAML cases was 23, reported in the article by Chai et al., 2024 [13]), which corresponds to the low overall prevalence of this subtype in the population. Due to the insufficient sample size, the data groups in several articles were not fully balanced for tumor size, sex, and age.

In a recent 2025 publication by Han J. H. et al. [14] on the differential diagnosis of SRMs using radiomic analysis, the authors attempted to address this issue by combining both AML subtypes – both lpAML and AML with high fat content (the quantitative ratio of subtypes in the overall sample was not provided, and sufficient attention was not paid to differentiating specifically lpAML), which could also affect the diagnostic accuracy of the results. Therefore, this publication was not included in the systematic review.

On the other hand, in 3 publications included in the systematic review, the patient sample contained various subtypes of renal cell carcinoma (clear cell, papillary, chromophobe), however, clear cell RCC numerically predominated, which could lead to bias in the models.

In 5 publications included in the systematic review, two-dimensional segmentation was used and 2D features were extracted, which may provide less information about the tumor compared to three-dimensional features; moreover, manual delineation of regions of interest was required.

Furthermore, this work has limitations characteristic of systematic reviews. The search was limited to English-language works, which likely somewhat reduced the number of identified studies. Data imbalance was observed in all studies. All this allowed us to perform only a qualitative synthesis using descriptive statistics, rather than a full meta-analysis.

Nevertheless, this study identified the main problematic points in current radiomics and the future direction of research in this area.

**Conclusions**

When considering the differential diagnosis of benign and malignant small renal masses (SRMs), it is important to note that one in five masses suspected of being cancer turns out to be benign [15]. This is one reason why radiomics has the potential to improve the preoperative detection of benign tumors.

However, the current ability to correctly determine stable radiomic features is limited: the reviewed publications demonstrate a wide variety of methodologies for image acquisition, post-processing, as well as feature extraction and statistical processing, which inevitably results in low reproducibility of radiomic features across studies.

According to previous research, radiomic analysis does not allow differentiation of lpAML from RCC with 100% specificity [16]. Consequently, analysis based solely on radiomics is insufficient for diagnosis.

In recent studies included in the review, integrated models based on both radiomic analysis data and general clinical and/or radiological data (radiomic nomograms) are increasingly common. Such a model can combine standard (conventional) CT image parameters (e.g., visual characteristics assessed by a radiologist) and radiomic features. In all reviewed publications, the AUC values for the radiomic nomogram surpassed those of both the isolated clinical model and the isolated radiomic model. This suggests that radiomic features complement traditional clinical and imaging characteristics.

In addition to radiomic models, machine learning models are being actively integrated into the study of renal tumors (several included studies used multiple algorithms simultaneously), which also demonstrate higher predictive performance. The application of deep learning can bring innovation to renal tumor research, and its combination with traditional methods, such as radiomics, may yield even better results in clinical practice.

Future research might consider integrating other imaging features and clinical factors to further enhance the effectiveness of radiomic nomograms. These studies additionally demonstrate that models built using multiple parameters possess higher predictive efficacy. The combined use of traditional imaging features and radiomics methods improves the ability to discriminate between types of small renal masses, providing substantial support for developing more precise patient treatment plans.

**Final Conclusion**

The use of non-invasive radiomic models for clinical decision-making holds significant potential. However, thorough validation of the obtained results is crucial before widespread implementation into clinical practice.

This study demonstrates to readers the need to improve the reproducibility of radiomic parameters and to identify a core set of robust radiomic parameters for implementing the method in clinical practice, which is only possible through standardization of radiomics methods and the creation of an open radiomic database. Further research is necessary before radiomics can become part of routine practice.

However, for the implementation of CT and MRI radiomics into clinical practice, not only standardization of image acquisition, processing, and analysis across institutions is needed, but also increased sample sizes in a multicenter format, prospective study designs, and external validation to create more reliable predictive models.

About the authors

Daria E. Zolotova

Lomonosov Moscow State University, Medical Research and Educational Institute (MREI MSU), Faculty of Fundamental Medicine

Email: darachich@yandex.ru
ORCID iD: 0009-0007-9567-3513
SPIN-code: 4907-7740

Russian Federation, Lomonosovsky Avenue, 27-10, Moscow

Elena A. Mershina

Lomonosov Moscow State University, Medical Research and Educational Institute (MREI MSU), Faculty of Fundamental Medicine

Email: elena_mershina@mail.ru
ORCID iD: 0000-0002-1266-4926
SPIN-code: 6897-9641

MD, Cand. Sci. (Med), Associate Professor

Russian Federation, Lomonosovsky Avenue, 27-10, Moscow

Lyudmila Andrianovna Makovskaya

Lomonosov Moscow State University, Medical Research and Educational Institute (MREI MSU), Faculty of Fundamental Medicine

Author for correspondence.
Email: l.a.makovskaya@yandex.ru
ORCID iD: 0000-0001-9127-7539
SPIN-code: 9668-2229

Врач отделения рентгенодиагностики с кабинетами МРТ и КТ Университетской клиники.

Russian Federation, Lomonosovsky Avenue, 27-10, Moscow

References

Supplementary files