Modified clear cell likelihood score and a new CAT score in the assessment of indeterminate small renal masses

Abstract

Purpose

In recent decades, small renal masses (SRMs) have become common incidental findings in cross-sectional studies; however, a widely implemented approach for the characterization of SRMs is still lacking. Thus, in this study, we aimed to explore the diagnostic performance of existing algorithms and to propose – as a conceptual framework rather than a clinically verified tool – a new simple radiological scale for estimating the probability of malignancy in SRMs.

Methods

Patients with indeterminate solid SRMs (N = 50), discovered using magnetic resonance imaging (MRI) between 2012 and 2023 were included. In 38 cases, the final diagnosis was based on histopathology, while in 12 cases it relied on regression or lack of progression during follow-up. Modified versions of the clear cell likelihood score (ccLS) were calculated, and its diagnostic performance was assessed. Moreover, we analyzed the newly created score, which consisted of selected MRI and clinical features. All of our modified scales used a Likert score for the likelihood of clear cell renal cell carcinoma (ccRCC).

Results

Based on the results of our statistical analyses, we modified the ccLS by adding T1 SI ratio < 0.73, arterial to delayed ratio (ADER) > 0.99, and smoking as independent predictors of ccRCC. We created a new scale, the CAT score, which combined hyperintensity in the Corticomedullary phase, ADER > 0.99, and TI SI ratio < 0.73 (with 1 point being assigned to each of the above-mentioned MRI parameters). In our results, the best diagnostic accuracy was observed for a CAT score ≥ 2, with a sensitivity of 73.9% (51.6–89.8%), a specificity of 77.8% (57.7–91.4%), and an accuracy of 76.0% (61.8–86.9%). Univariate logistic regression analyses demonstrated that all scales created by our group were significant predictors of ccRCC. Importantly, they showed a better predictive ability than the standard ccLS score.

Conclusions

The CAT score appears to improve the prediction of ccRCC compared with both standard and modified versions of the ccLS and may serve as a potential aid in the routine assessment of indeterminate SRMs. Nonetheless, this study should be regarded as a preliminary, proof-of-concept analysis rather than a definitive model-development study. The main limitation of our study is its small, single-center cohort, which limits the statistical power and robustness of model development. Moreover, a substantial proportion of benign diagnoses based only on radiological follow-up rather than histopathology. Therefore, before clinical implementation, the CAT score requires prospective validation in larger, independent, multicenter cohorts.

Introduction

In recent decades, small renal masses (SRMs) have become common incidental findings in cross-sectional studies, and proper determination of this type of lesion based on imaging features has proven to be notoriously challenging (Sanchez et al. 2018; Sebastia et al. 2020). Computed tomography with contrast enhancement is an excellent modality for the initial characterization of most SRMs; however, the conventional approach to computed tomography interpretation offers only a limited capability to accurately predict the histopathological type of lesions (Krishna et al. 2017). Magnetic resonance imaging allows for further assessment of SRMs due to its superior soft-tissue contrast, the ability to employ chemical shift imaging and diffusion-weighted imaging, as well as dynamic imaging after administration of gadolinium-based agents (Pedrosa and Cadeddu 2022).

Although 80% of SRMs are malignant, most of them exhibit indolent oncologic behavior with only a minimal risk of rapid tumor growth and metastases (Sanchez et al. 2018). Contemporary data show that for most SRMs ≤ 4 cm, surgery or ablation does not improve cancer-specific survival compared with active surveillance, particularly in older patients or those with significant comorbidities, in whom competing risks of death predominate and nephron loss may worsen cardiovascular outcomes (Patel et al. 2012; Pierorazio et al. 2015; Uzosike et al. 2018; Cooperberg et al. 2008; Hollingsworth et al. 2006). Thus, the American Urological Association (AUA) guidelines note that for a clinical T1a mass active surveillance with potential for delayed intervention should be discussed as an option for patients under certain conditions. The guideline stresses individualized, shared-decision-making: patient’s comorbidities, life expectancy, tumor characteristics (size, growth potential, imaging appearance) should all be considered (Schieda et al. 2022a, b; Campbell et al. 2021). However, per current EAU (2025) guidelines, partial nephrectomy remains the standard of care and is strongly recommended for T1a tumors in fit patients when technically feasible, offering equivalent oncological outcomes to radical nephrectomy while better preserving renal function; ablation may be considered as an alternative for frail individuals with small masses (Campbell et al. 2021).

In the case of early-detected clear cell renal cell carcinoma (ccRCC, the most common aggressive type of kidney cancer best managed with prompt surgical treatment), active surveillance may delay definitive treatment, leading to adverse outcomes and compromising a patient’s chance for long-term survival (Hao et al. 2023a, b; Smaldone et al. 2012; Vazquez et al. 2024).

Despite advances in imaging, a standardized and widely adopted approach for the characterization of SRMs is still lacking. The clear cell likelihood score (ccLS) algorithm was proposed to provide guidance for the prediction of ccRCC risk in multiparametric MRI studies (Pedrosa and Cadeddu 2022; Hao et al. 2023a, b; Vazquez et al. 2024; Chen et al. 2024; Shetty et al. 2023; Pedrosa 2023; Tse 2022; Shetty 2022; Rasmussen et al. 2022; Mileto and Potretzke 2022; Dunn et al. 2022; Cui et al. 2022; Diaz de Leon et al. 2019). However, the ccLS was assessed in several institutions and reports show only a moderate diagnostic accuracy, similar to that of radiologist assessment (Pedrosa and Cadeddu 2022; Hao et al. 2023a, b; Tse 2022; Mileto and Potretzke 2022; Dunn et al. 2022; Tian et al. 2022; Canvasser et al. 2017; Schieda et al. 2022a, b).

Recent evidence further highlights the clinical relevance of accurately characterizing SRMs, as both surgical decision-making and tumor biology critically influence patient outcomes. Studies on robotic and nephron-sparing surgery show that surgical technique—particularly the use of tumor enucleation—can reduce perioperative complications, preserve renal function, and may be associated with low rates of positive surgical margins, while standard resection, older age, upper-pole location, and high-grade histology increase the risk of incomplete excision (Bertolo et al. 2023; Schiavina et al. 2015). At the same time, advances in molecular oncology reveal that dysregulated microRNAs contribute to RCC development, progression, and treatment resistance, underscoring the need for improved diagnostic tools capable of distinguishing aggressive ccRCC from indolent lesions (Napolitano et al. 2023). These findings emphasize that precise, reliable preoperative characterization of SRMs is essential to optimize treatment selection, avoid unnecessary interventions, and ensure timely management of biologically aggressive tumors.

Therefore, the aims of this study were to evaluate the diagnostic accuracy of the modified ccLS algorithm and to develop and internally assess a new scoring system that may improve ccRCC prediction in SRMs compared with the standard or modified ccLS models.

Methodology

Studied population

Our retrospective, single-center study included 50 patients with indeterminate SRMs discovered using MRI imaging. In 38 cases, the final diagnosis was determined using histopathological results as the standard reference, while in 12 cases, it was based on regression or a complete absence of progression observed on follow-up (minimum 36 months, mean 45.58 ± 9 months). Only SRMs without obvious malignant features (e.g. large, infiltrative tumors) or obvious benign features (e.g. fat-rich AMLs) were included in the analysis. Patients with incomplete clinical data or those for whom a definitive diagnosis was not possible due to a short follow-up period were excluded from the study. Flowchart illustrating the patient selection process is depicted on Fig. 1. This study was performed in accordance with the Declaration of Helsinki, was approved by the Ethics Committee, and the requirement for informed patient consent was waived (OIL/KBL/51/2023).

Fig. 1.

Flowchart illustrating the patient selection process. Abbreviations: AMLs – angiomyolipomas, MRI – magnetic resonance imaging.

Magnetic resonance imaging studies

Magnetic resonance imaging studies were performed using 1.5T and 3.0T scanners in the Department of Diagnostic Imaging in the University Hospital of Cracow. With only minimal variability in the study protocols, all studies included the necessary sequences for lesion assessment according to ccLS v2.0: axial and coronal 2D T2w single shot acquisitions, axial 2D T1w gradient echo in and out of phase for chemical shift imaging, pre- and (dynamic) post-contrast 3D T1w SPGR with fat saturation including delayed scans, and diffusion-weighted imaging. All MRI examinations were retrospectively reviewed by two radiologists (T.B. and J.D.J.) who were blinded to the final diagnosis. The following MRI parameters were used for our analysis: T1 signal intensity (SI) ratio (measured as the quotient of tumor SI to renal cortex SI), hyperintensity in the corticomedullary phase, and arterial to delayed ratio (ADER, obtained as the quotient of difference between SI in the corticomedullary phase and pre-contrast images to the difference in SI on delayed phase and pre-contrast images).

In accordance with the protocols delineated in the ccLS for multiparametric MRI evaluation of renal masses, we adopted analogous criteria for T1-weighted signal intensity measurements as those recommended for T2-weighted sequence assessments. Specifically, T1-weighted signal intensity was measured within the enhancing segments of the renal masses, following meticulous evaluation of the delayed post-contrast phase to localize viable tissue and preclude non-enhancing regions, including those exhibiting hemorrhage, necrosis, or cystic degeneration. This approach involved placement of a region of interest (ROI) in the tumor area demonstrating maximal contrast enhancement, thereby ensuring focus on viable, enhancing components and mitigating potential variability in diagnostic interpretation.

The clear cell likelihood score, its modifications, and other new scores

First, we calculated the ccLS score for our population as a standardized framework for categorizing SRMs (Shetty et al. 2023). Next, we modified the classic ccLS algorithm in consecutive steps in order to optimize it. Moreover, we created a new score consisting of different MRI and clinical features. According to our previous results (Blachura et al. 2024), we used the following cut-off points for the MRI parameters: ADER > 0.99 and TI SI ratio < 0.73. All of our modified scales used a Likert score for the likelihood of ccRCC.

Statistical analysis

Continuous variables are presented as means ± standard deviations or medians and interquartile ranges (IQR). They were compared between the study groups using the Student’s t-test or the Mann-Whitney test, as appropriate. Categorical variables are presented as numbers and percentages and were evaluated by the Pearson χ2 test or Fisher’s exact test. The McNemar test was used to evaluate the agreement between the newly tested criteria and diagnosis of ccRCC based on histopathology (ccRCC vs. ‘all other’ histologies). Specificity, sensitivity, positive predictive value, negative predictive value, accuracy, and negative likelihood ratio were calculated for each tested criterion. Receiver operating characteristic (ROC) curve analysis was performed to compare the diagnostic performance of the CAT score and the classic ccLS in predicting ccRCC. A p-value of 0.05 or less was considered statistically significant. Statistical analyses were performed using IBM SPSS Statistics (version 24, IBM Corp., Armonk, NY, USA) and StataNow/BE (version 19.5, StataCorp LLC, College Station, TX, USA). Confidence intervals (CI) were calculated using MEDCALC (free statistical calculators).

Results

Patient characteristics

A total of 50 subjects with SRMs (27 females, 23 males; median age 61.5 [IQR: 52.3–68.3] years) were included in the study. There were significantly more smokers in the ccRCC group than in the non-ccRCC group (14 [60.9%] vs. 8 [29.6%], p = 0.03). Apart from this, the ccRCC and non-ccRCC groups did not differ in terms of main risk factors (male sex, symptoms, hypertension, family history of neoplasm) related to ccRCC development, as reported previously (Blachura et al. 2024).

Associations between selected MRI parameters and final diagnoses of SRMs

We found that the numbers of cases with ADER > 0.99 and TI SI ratio < 0.73 were significantly different between the ccRCC and non-ccRCC groups (65% vs. 33%, p = 0.03; 39% vs. 4%, p = 0.003; respectively). The McNemar test demonstrated that only ADER > 0.99 and hyperintensity in the corticomedullary phase were in agreement with the final diagnosis (Table 1).

Table 1.

Differences in selected MRI parameters in the detection of ccRCC

MRI parameter	ccRCC (n = 23)		non-ccRCC (n = 27)		p-value	McNemar test
	TP	FN	FP	TN
ADER > 1.5	1 (4%)	22 (96%)	24 (89%)	3 (11%)	0.38	< 0.001
ADER > 0.99	15 (65%)	8 (35%)	18 (67%)	9 (33%)	0.03	1
T1 SI ratio < 0.73	9 (39%)	14 (61%)	26 (96%)	1 (4%)	0.003	< 0.001
Hyperintensity in cortico-medullary phase	17 (74%)	6 (26%)	13 (48%)	14 (52%)	0.11	0.12

Abbreviations: ADER - arterial to delayed enhancement ratio; ccRCC – clear cell renal cell carcinoma; FP – false positive; FN – false negative; MRI – magnetic resonance imaging; non-ccRCC – other than clear cell renal cell carcinoma; TP -true positive; TN – true negative; SI – signal intensity

Diagnostic accuracy of the modified ccLS and other new scales in the prediction of ccRCC

We modified the ccLS based on the results of our previous statistical analyses. First, we modified the ccLS by adding the T1 SI ratio < 0.73 as an independent predictor of ccRCC (Tables 2 and 3), which resulted in a sensitivity of 47.8% (26.8–69.4%), a specificity of 96.3% (81.0-99.9%), and an accuracy of 74.0% (59.7–85.4%).

Table 2.

Different scores in patients with ccRCC vs. patients with non-ccRCC etiology

Score	ccRCC (n = 23)		non-ccRCC (n = 27)		p-value
	TP	FN	TN	FP
ADER + T1
= 0	4 (17%)	19 (83%)	10 (37%)	17 (63%)	0.001
= 1	14 (61%)	9 (39%)	17 (63%)	10 (37%)	0.09
= 2	5 (22%)	18 (78%)	27 (100%)	0 (0%)	0.02
ADER + T1 + smoking
= 0	1 (4%)	22 (96%)	16 (59%)	11 (41%)	0.003
= 1	9 (39%)	14 (61%)	13 (48%)	14 (52%)	0.37
= 2	10 (43%)	13 (57%)	25 (93%)	2 (7%)	0.003
= 3	3 (13%)	20 (87%)	27 (100%)	0 (0%)	0.09
≥ 2	13 (57%)	10 (43%)	25 (93%)	2 (7%)	< 0.001
CAT score
= 1	3 (13%)	20 (87%)	15 (56%)	12 (44%)	0.02
= 2	13 (57%)	10 (43%)	21 (78%)	6 (22%)	0.01
= 3	4 (17%)	19 (83%)	27 (100%)	0 (0%)	0.04
≥ 2	17 (74%)	6 (26%)	21 (78%)	6 (22%)	< 0.001
ccLS + T1 SI ratio < 0.73
≥ 5	11 (48%)	12 (52%)	26 (96%)	1 (4%)	< 0.001
ccLS + T1 SI ratio < 0.73 + ADER > 0.99
≥ 5	15 (65%)	8 (35%)	23 (85%)	4 (15%)	< 0.001
ccLS + T1 SI ratio < 0.73 + ADER > 0.99 + smoking
≥ 6	14 (61%)	9 (39%)	24 (89%)	3 (11%)	< 0.001

Abbreviations: ADER - arterial to delayed enhancement ratio; ccLS – clear cell likelihood score; ccRCC – clear cell renal cell carcinoma; FP – false positive; FN – false negative; MRI – magnetic resonance imaging; non-ccRCC – other than clear cell renal cell carcinoma; TP -true positive; TN – true negative; SI – signal intensity

Table 3.

Sensitivity, specificity, accuracy, PLR, NLR, PPV, and NPV of the modified ccLS

Modified ccLS	Sensitivity	Specificity	Accuracy	PLR	NLR	PPV	NPV
ccLS + T1 SI ratio ≥ 5	47.8 (26.8–69.4)%	96.3 (81.0-99.9)%	74.0 (59.7–85.4)%	12.9 (1.8–92.6)	0.5 (0.4–0.8)	91.7 (60.5–98.8)%	74.0 (59.7–85.4)%
ccLS + T1 SI ratio + ADER ≥ 5	65.2 (42.7–83.6)%	85.2 (66.3–95.8)%	76.0 (61.7–83.7%)	4.4 (1.7–11.4)	0.4 (0.2–0.7)	79.0 (59.1–90.7)%	74.2 (61.7–83.7)%
ccLS + T1 SI ratio + ADER + smoking ≥ 6	60.9 (38.5–80.3)%	88.9 (70.8–97.7)%	76.0 (61.8–86.9)%	5.5 (1.8–16.7)	0.4 90.3–0.8)	82.4 (60.5–93.4)%	72.7 (61.2–81.9)%

Abbreviations: ADER - arterial to delayed enhancement ratio; ccLS – clear cell likelihood score; SI – signal intensity; PLR – positive likelihood ratio; NLR – negative likelihood ratio; PPV – positive predictive value; NPV – negative predictive value

In the second step, we added ADER > 0.99 and observed that the sensitivity increased to 65.2% (42.7–83.6%), specificity decreased to 85.2% (66.3–95.8%), and accuracy increased to 76.0% (61.7–83.7%).

In the final step, we added a clinical feature: smoking. This slightly increased specificity to 88.9% (70.8–97.7%) but reduced sensitivity to 60.9% (38.5–80.3%). Accuracy was similar, at 76.0% (61.8–86.9%).

Furthermore, we combined the TI SI ratio < 0.73, ADER > 0.99, and smoking to determine whether any of these combinations are better than the standard ccLS (Tables 2 and 4).

Table 4.

Sensitivity, specificity, accuracy, PLR, NLR, PPV, and NPV of the combined selected MRI and clinical features

Score	Sensitivity	Specificity	Accuracy	PLR	NLR	PPV	NPV
ADER + T1
= 0	17.4 (5.0–38.8)%	37.0 (19.4–57.6)%	28.0 (16.2–42.5)%	0.3 (0.1–0.7)	2.2 (1.2–3.8)	19.1 (8.4–37.5)%	34.5 (23/7–47.1)%
= 1	60.9 (38.5–80.3)%	63.0 (42.4–80.6)%	62.0 (47.2–75.4)%	1.6 (0.9–3.0)	0.6 (0.4–1.1)	58.3 (43.7–71.7)%	65.4 (51.3–77.2)%
= 2	21.7 (7.5–43.7)%	100.0 (87.2–100.0)%	64.0 (49.2–77.1)%	–	0.8 (0.6–1.0)	100.0 (47.8–100.0)%	60.0 (54.7–65.0)%
ADER + T1 + smoking
= 0	4.4 (0.1–22.0)%	59.3 (38.8–77.6)%	34.0 (21.2–48.8)%	0.1 (0.01–0.8)	1.6 (1.2–2.2)	8.3 (1.3–39.5)%	42.1 (34.5–50.2)%
= 1	57.6 (39.2–74.5)%	48.2 (28.7–68.1)%	53.3 (40.0–66.3)%	1.1 (0.7–1.8)	0.9 (0.5–1.5)	57.6 (46.0–68.4)%	48.2 (34.7–61.9)%
= 2	43.5 (23.2–65.5)%	92.6 (75.7–99.1)%	70.0 (55.4–82.1)%	5.9 (1.4–24.1)	0.6 (0.4–0.9)	83.3 (55.0–95.4)%	65.8 (57.0–73.7)%
= 3	13.0 (2.8–33.6)%	100.0 (87.2–100.0)%	60.0 (45.2–73.6)%	–	0.9 (0.7–1.0)	100.0 (29.2–100.0)%	57.5 (53.5–61.3)%
≥ 2	56.5 (34.5–76.8)%	92.6 (75.7–99.1)%	76.0 (61.8–86.9)%	7.6 (1.9–30.4)	0.5 (0.3–0.8)	86.7 (62.0–96.3)%	71.4 (60.8–80.1)%

Abbreviations: ADER - arterial to delayed enhancement ratio; PLR – positive likelihood ratio; NLR – negative likelihood ratio; PPV – positive predictive value; NPV – negative predictive value

Diagnostic accuracy of the novel CAT score in the prediction of CcRCC

We created a new score (CAT), which combined the following MRI parameters: hyperintensity in the Corticomedullary phase, ADER > 0.99, and TI SI ratio < 0.73, in which 1 point may be assigned for each parameter.

The methodology of the CAT score calculation is presented in Fig. 2.

Fig. 2.

The methodology of the CAT score calculation

We tested whether our new score improves prediction of ccRCC when compared to other approaches used for ccRCC assessment. In our results, the greatest diagnostic accuracy was observed for a CAT score ≥ 2, with a sensitivity of 73.9% (51.6–89.8%), a specificity of 77.8% (57.7–91.4%), and an accuracy of 76.0% (61.8–86.9%) (Table 5).

Table 5.

Sensitivity, specificity, accuracy, PLR, NLR, PPV, and NPV of the CAT score

CAT score	Sensitivity	Specificity	Accuracy	PLR	NLR	PPV	NPV
= 0	13.0 (2.8–33.6)%	66.7 (46.0–83.5)%	42.0 (28.2–56.8)%	0.4 (0.1–1.3)	1.3 (1.0–1.8)	25.0 (9.3–52.1)%	47.4 (39.8–55.1)%
= 1	13.0 (2.8–33.6)%	55.6 (35.3–74.5%)	36.0 (22.9–50.8)%	0.3 (0.1–0.9)	1.6 (1.1–2.3)	20.0 (7.4–43.8)%	42.9 (34.1–52.1)%
≤ 1	26.1 (10.2–48.4)%	22.2 (8.6–42.3)%	24.0 (13.1–38.2%)	0.3 (0.2–0.7)	3.3 (1.6–7.0)	22.2 (12.2–36.9)%	26.1 (14.3–42.7)%
= 2	56.5 (34.5–76.8)%	77.8 (57.7–91.4)%	68.0 (53.3–80.5)%	2.5 (1.2–5.6)	0.6 (0.3–0.9)	68.4 (49.5–82.7)%	67.7 (55.8–77.7)%
= 3	17.4 (5.0–38.8)%	100.0 (87.2–100.0)%	62.0 (47.2–75.4)%	–	0.8 (0.7–1.0)	100.0 (39.8–100.0)%	58.7 (54.1–63.2)%
≥ 2	73.9 (51.6–89.8)%	77.8 (57.7–91.4)%	76.0 (61.8–86.9)%	3.3 (1.6–7.0)	0.3 (0.2–0.7)	73.9 (57.3–85.7)%	77.8 (63.1–87.8)%

Abbreviations: PLR – positive likelihood ratio; NLR – negative likelihood ratio; PPV – positive predictive value; NPV – negative predictive value

Diagnostic accuracy of ccLS and CAT score used together in the prediction of ccRCC

Using a combination of CAT score ≥ 2 and ccLS ≥ 4, we observed that 12 cases were classified as true positive and 23 as true negative (p = 0.005), with a sensitivity of 52.2% (30.6–73.2%), a specificity of 85.2% (66.3–95.8%), and an accuracy of 70.0% (55.4–82.1%).

The classic ccLS, its modifications, and CAT score in the prediction of ccRCC

Univariate logistic regression analyses demonstrated that all of the scales created by us were significant predictors of ccRCC (Table 6). Importantly, they showed a better predictive ability than the standard ccLS score.

Table 6.

Univariate logistic regression in predicting an occurrence of ccRCC

Parameter	OR	95% CI	p-value
ccLS ≥ 4	2.3	0.7–7.0	0.16
ccLS = 5	9.2	1.0–83.1	0.049
ccLS + T1 SI ratio ≥ 5	23.8	2.8–206.3	0.004
ccLS + T1 SI ratio + ADER ≥ 6	10.8	2.8–42.2	< 0.001
ccLS + T1 SI ratio + ADER + smoking ≥ 7	12.4	2.9–53.8	< 0.001
CAT score ≥ 2	9.9	2.7–36.4	< 0.001
CAT score ≥ 2 and ccLS ≥ 4	6.3	1.6–24.0	0.007

Abbreviations: ADER - arterial to delayed enhancement ratio; ccLS – clear cell likelihood score; CI - confidence interval; OR – odds ratio

McNemar’s test demonstrated a statistically significant difference in classification accuracy between the CAT score and the ccLS model (p < 0.001), indicating that the two models do not perform equivalently on the same set of cases. Specifically, the CAT score more frequently produced correct classifications in cases where the ccLS misclassified the lesion, supporting its potential diagnostic advantage.

Receiver operating characteristic curve analysis was performed to compare the diagnostic performance of the CAT score and the classic ccLS in predicting ccRCC. The resulting ROC curves for both models are presented in Fig. 3. The CAT score achieved an area under the curve (AUC) of 0.7705 (95% CI: 0.6402–0.9008), indicating good discriminative ability. The AUC for the ccLS was 0.6868 (95% CI: 0.5468–0.8268), corresponding to moderate diagnostic performance.

Fig. 3.

Receiver operating characteristic curves illustrating the diagnostic performance of the CAT score and the ccLS in differentiating clear cell renal cell carcinoma from other small renal masses

Although the CAT score demonstrated a greater AUC than the ccLS, the difference between the two was not statistically significant (χ²(1) = 1.47, p = 0.23, DeLong test), suggesting comparable overall diagnostic accuracy within this cohort. Nevertheless, the greater point estimate of the AUC for the CAT score supports its potential as a practical and effective tool for differentiating ccRCC from other histopathological subtypes among SRMs.

Using the Youden index to determine the optimal cut-off, the best threshold for the CAT score was ≥ 2 (empirical cut-point = 1.5), providing the best balance between sensitivity (0.74) and specificity (0.78). For the ccLS, the optimal cut-off was ≥ 4 (empirical cut-point = 3.5), corresponding to a sensitivity of 0.61 and a specificity of 0.59. These results confirm that the CAT score offers better overall discriminatory performance than the classic ccLS in differentiating ccRCC from other histopathological subtypes of SRMs.

Discussion

Many reports have shown the ability of various imaging techniques to predict the histological outcomes of SRMs (Galmiche et al. 2017; Lee-Felker et al. 2014; Rosenkrantz et al. 2010). However, there is still no standardized approach for the characterization of indeterminate SRMs on MRI, as the diagnostic performance of ccLS is variable (Tian et al. 2022). In our population, the classic ccLS ≥ 4 had an accuracy of 60.0%, a sensitivity of 59.3% and a specificity of 60.9%. After modification of the classic ccLS by adding new parameters (T1 SI ratio < 0.73 and ADER > 0.99), we observed better sensitivity, specificity, and accuracy of this modified scale (for modified ccLS ≥ 5: 65.2%, 85.2%, 76.0%, respectively). This suggests that using a T1 SI ratio < 0.73 and ADER > 0.99 could strengthen the diagnostic performance of the ccLS.

The concept of combining straightforward imaging and clinical features into a simple, reproducible diagnostic score aligns with approaches used in other medical specialties. For example, the CAR₂E₂ score, which integrates clinical, radiographic, and electrocardiographic parameters for left ventricular hypertrophy screening, demonstrated that such composite models can effectively enhance diagnostic performance in routine clinical settings (Matusik et al. 2022). Importantly, according to the initial results, we discovered that our new CAT score, based on hyperintensity in the corticomedullary phase, ADER > 0.99, and T1 SI ratio < 0.73, has a better diagnostic accuracy than both the classic and modified ccLS. For CAT score ≥ 2, we observed a sensitivity of 73.9%, a specificity of 77.8%, and an accuracy of 76.0%. This demonstrates that the above parameters may have significant importance in the differential diagnosis of SRMs, especially in ccRCC diagnosis. The results of our analysis are consistent with the existing literature. Generally, ccRCC presents as intense enhancement in the corticomedullary phase, and moderate “washout” when compared to fat-poor angiomyolipomas or other non-malignant or indolent lesions (i.e. papillary RCC) (Krishna et al. 2017) (Pedrosa and Cadeddu 2022) (Sasiwimonphan et al. 2012). Currently, there is a paucity of data regarding values of the T1 SI ratio in ccRCC; however, a lower T1 SI ratio is associated with invasive lesions (Wang et al. 2020). It has been demonstrated that a T1 SI ratio ≥ 2.15 is a highly specific predictor for non-enhancement (Le et al. 2015). In another study, it was demonstrated that SRMs with a T1 SI ratio > 1.6 may predict benign lesions (Davarpanah et al. 2016). Generally, a T1 SI ratio > 1 means that the lesion is hypersignal, i.e. with macroscopic fat if T1 is without saturation or a cyst with dense contents/bleeding.

Renal cell carcinoma, similar to other cancers, is often multifactorial in origin, and a combination of various factors may work together to increase the risk. Smoking, arterial hypertension, and obesity are risk factors thought to account for roughly 50% of ccRCC (Haggstrom et al. 2013). In our study, smoking was the only clinical parameter having a statistically significant association with ccRCC risk (60.9% in the ccRCC group, 29.6% in the non-RCC group). However, using this parameter did not significantly improve the diagnostic performance of our modified scales in the prediction of ccRCC. Adding smoking as a parameter slightly increases the specificity, but reduces the sensitivity of the scale. This confirms that the preliminary characterization of SRMs based only on imaging features is possible.

In recent years, several artificial intelligence (AI)-driven approaches, including radiomics and deep learning models, have been proposed for the characterization of SRMs (Rosenkrantz et al. 2010; Matusik et al. 2022). These models frequently demonstrate high diagnostic performance, often surpassing that of traditional radiologist-based assessments in differentiating RCC subtypes. For instance, recent studies have reported pooled sensitivities and specificities of 85% and 76%, respectively, for internally validated AI models, with external validations showing even higher specificity (90%) and slightly lower sensitivity (80%) (Sasiwimonphan et al. 2012). Compared to the CAT score, which in our study achieved an estimated sensitivity of 73.9% and specificity of 77.8%, these AI models may offer improved diagnostic accuracy. However, it is important to highlight that direct, head-to-head comparisons between AI algorithms and radiologist-driven scoring systems (including both ccLS and CAT score) remain scarce, with most available comparisons focusing solely on ccLS. While the CAT score provides a practical, interpretable tool for radiologists that builds upon standard imaging features, AI-based models—particularly those incorporating radiomics or deep learning—may represent a paradigm shift toward precision diagnostics in renal oncology. Importantly, the integration of AI with radiologist expertise, for instance, through hybrid models combining algorithmic frameworks like the CAT score or ccLS with AI-generated insights, could potentially yield synergistic benefits. Future prospective studies are needed to validate such approaches and determine their feasibility, generalizability, and clinical utility in real-world settings.

Compared with emerging AI- and radiomics-based models, the CAT score offers several practical advantages. While AI approaches may ultimately achieve higher predictive accuracy, they often require advanced computational infrastructure, proprietary software, and may function as “black-box” systems with limited interpretability. In contrast, the CAT score is a transparent, rule-based system derived solely from standardized imaging features readily observable by radiologists. Each component—such as signal intensity ratios and enhancement patterns—is explicitly defined and can be independently verified, ensuring that the entire decision-making process remains understandable and reproducible. This interpretability may enhance diagnostic confidence and facilitate integration into routine clinical workflows, particularly in settings without access to AI-based tools. However, a direct, head-to-head comparison between the CAT score and radiologist assessment was not performed in this study. Thus, the main advantage of the CAT score over AI/radiomics-based models lies not in superior raw performance, but in its interpretability, transparency, and straightforward clinical applicability.

This study has several limitations that should be acknowledged. First, its retrospective, single-center design may introduce patient selection bias and limits the generalizability of the findings. Second, the relatively small cohort—comprising 50 renal masses, of which 23 were confirmed as ccRCC—reduces the statistical power and increases the risk of model overfitting. Although histopathological verification served as the gold standard for most cases, 12 lesions were diagnosed as benign based solely on stable clinical and imaging follow-up over a 3-year period without evidence of progression, which may introduce a degree of diagnostic uncertainty. Additionally, we observed a trend toward larger tumor size in the ccRCC group, which likely reflects a biological and radiological bias, as larger lesions tend to display more distinctive imaging features that facilitate classification. Reproducibility remains insufficiently addressed in this study. Although imaging assessments were performed by radiologists with subspecialty experience, no inter-reader agreement metrics (such as kappa or intraclass correlation coefficients) were calculated for the three MRI parameters that constitute the backbone of the CAT score—T1 signal intensity ratio, ADER, and corticomedullary hyperintensity. Even though the inter-reader reliability for the final ccLS categories in our cohort was good (κ = 0.63), the absence of component-level reproducibility data limits the evaluation of the CAT score’s robustness, particularly when applied by different readers in routine clinical practice. Variability in these foundational measurements may influence the resulting CAT score, especially in borderline cases, and could reduce the generalizability of the reported diagnostic performance. Scanner heterogeneity (1.5 T vs. 3 T) represents another relevant source of bias. SI-based ratios may shift across field strengths or minor protocol variations, even when similar acquisition parameters are used. Although all examinations in this study followed standardized institutional protocols, differences in image characteristics between scanners cannot be fully eliminated. The use of ratios instead of absolute signal intensities may reduce field-strength–related variability, but small shifts in these measurements remain possible and could affect individual components of the CAT score. Future work should incorporate formal reproducibility analyses. Finally, the CAT score represents a preliminary, exploratory model derived from a limited dataset, providing early evidence that it may enhance the prediction of ccRCC compared with standard or modified ccLS algorithms. However, these results should be interpreted as hypothesis-generating. Prospective validation in larger, independent, multicenter cohorts is essential to confirm the model’s reliability, generalizability, and potential clinical utility before clinical implementation.

Conclusions

The CAT score demonstrated a potential to enhance the discrimination of ccRCC compared with both the standard and modified ccLS, suggesting its possible value in evaluating indeterminate SRMs. However, these findings should be interpreted within the context of an early proof-of-concept study rather than a fully developed predictive model. The limited, single-center cohort substantially reduces statistical power and the stability of the reported performance metrics, and many benign lesions were classified without histopathological confirmation, relying solely on imaging follow-up. Consequently, the CAT score should undergo rigorous prospective validation in larger, independent, multicenter cohorts before it can be considered for clinical application.

Author contributions

Tomasz Blachura, MD – conceptualization, data curation, formal analysis, manuscript writing (original draft).Julia Radzikowska, MD – investigation, data collection, manuscript editing.Patrycja S. Matusik, MD, PhD – supervision, conceptualization, methodology design, manuscript review and editing.Aleksander Kowal, MD – data collection, image analysis, visualization.Jarosław D. Jarczewski, MD – data collection, manuscript editing.Łukasz Skiba, MD – data collection, manuscript editing.Tadeusz J. Popiela, MD, PhD – supervision, critical revision of the manuscript.Robert Chrzan, MD, PhD – suprvision, conceptualization, critical revision of the manuscript.All authors have read and approved the final version of the manuscript.

Data availability

All data are available in the manuscript.

Declarations

Conflict of interests

The authors declare no competing interests.