Renal cell carcinoma (RCC) is among the most commonly diagnosed forms of cancer among both men and women in the United States with 2% growth in incidence [1, 2]. Up to 70% of renal neoplasms are detected incidentally at imaging, and all solid enhancing renal masses without macroscopic fat are presumed to be malignant. However, approximately 20% of clinical T1a small renal masses (<4 cm) at nephrectomy and 30% of lesions at biopsy are benign, most commonly oncocytoma and fat-poor angiomyolipoma (fp-AML). Thus, a significant percentage of patients with incidentally detected renal masses are subjected to the risks of unnecessary procedures [3–7].

While these benign lesions, and more indolent RCCs such as chromophobe RCC (chRCC) and papillary RCC (pRCC), may undergo active surveillance, clear cell RCC (ccRCC) is more likely to be treated with resection or ablation because it tends to be more aggressive with higher metastatic potential [7–10]. Although a percutaneous biopsy to confirm renal mass histology has become broadly accepted, it may have non-diagnostic rate of 10–20% and a relatively poor negative predictive value [11, 12]. Multiphasic computed tomography (CT) or magnetic resonance (MR) imaging has become the standard for characterizing masses in the liver, lung, and brain, but has not yet been established in the kidney [11, 13–15]. Thus, an imaging-based method to differentiate malignant from benign renal lesions, and also to distinguish the more aggressive ccRCC from less aggressive malignant subtypes, would be of great clinical value, particularly in older patients with comorbidities in whom a conservative approach would be favored.

Several studies have attempted to characterize renal masses on imaging by correlating unique multiphasic patterns of contrast attenuation with the histology of RCC subtypes and benign lesions on CT and MR. These patterns are identified by visual assessment of peak lesion attenuation and manual placement of a region of interest (ROI) within the most enhancing portion of the lesion and background cortex over multiple phases, often without the use of a standardized four-phase renal protocol [16–28]. This technique is inherently operator dependent and subjective, with variable results reported. A reproducible and robust method to differentiate renal lesions on imaging is desired to minimize subjectivity and ensure widespread applicability. The purpose of our study was to evaluate the performance of a novel, computer-aided diagnostic (CAD) algorithm that detects peak lesion attenuation relative to renal parenchyma to differentiate ccRCC from chRCC, pRCC, oncocytoma, and fp-AML on multiphasic CT and compare CAD performance with the manual method.

Materials and methods

Patients and lesions

With institutional review board approval for this 1996 Health Insurance Portability and Accountability Act compliant retrospective study and a waiver of informed consent, we queried our institution’s pathology database for histologically proven cases of RCC, oncocytoma, and fp-AML from January 2000 to 2016. We searched our institution’s medical records to identify the date of surgery or biopsy, and our image archiving system (Centricity; GE Medical Systems, Milwaukee, WI) to identify patients in which a preoperative MDCT with a four-phase renal protocol had been acquired. The importance of our study was in the context of patients for whom quantitative data would be most beneficial in assisting clinicians with diagnosis at the time of initial presentation, management during active surveillance or triaging a patient to a minimally invasive therapy. Therefore, our query was limited to patients with pathological stage T1 renal masses, as the course of treatment for a patient with a renal mass >7 cm stage T2–T4 would most likely be surgical intervention. To avoid potential clustering effects, only patients with a solitary tumor were included. Our final cohort consisted of 200 consecutive lesions in 200 unique patients (Fig. 1).

Fig. 1
figure 1

Patient flow diagram. CT computed tomography, RCC renal cell carcinoma

Full size image

CT acquisition and four-phase renal mass protocol

All patients had a preoperative MDCT at our institution with an intravenous (i.v.) contrast-enhanced four-phase renal mass protocol. This included an unenhanced scan and contrast-enhanced acquisitions during the corticomedullary, nephrographic, and excretory phases acquired from the lung bases to the iliac crests. Preoperative MDCT was performed on multiple scanners (4-MDCT, 16-MDCT, and 64-MDCT) from January 2000 to January 2016. Acquisition parameters were similar over the multiple generations of scanners: 120 kVp, 200–400 mA based on patient size and 3-mm scan collimation with 3-mm reconstruction intervals. Patients received a power injection of non-ionic i.v. contrast (iodixanol, Omnipaque 350, GE Healthcare), 35–45g iodine dosed to weight at a rate of 3 mL/s. Patients who weighed less than 45 kg, those who weighed 45–90 kg, and those who weighed >90 kg received 100 mL (35 g iodine), 125 mL (45 g iodine), and 150 mL (54 g iodine) of contrast material, respectively. A bolus tracking system (Care Bolus VB10, Siemens Healthcare or Smart-Prep, GE Healthcare) was used to determine the onset of imaging and a ROI was placed in the thoracoabdominal aorta junction, with a trigger set to begin at 150 HU. Images were acquired 55 s (corticomedullary), 120 s (nephrographic), and 8 min (excretory) after the threshold of 150 HU was reached.

Peak lesion attenuation: CAD method

Using in-house-developed U.S. FDA 510K cleared image processing software, the entire renal mass was contoured on axial images in each of the four phases, and was manually edited to ensure the contour was limited to the target lesion and excluded adjacent perirenal fat and normal renal parenchyma. We developed a CAD algorithm to automatically place an ROI in each lesion to quantify peak lesion attenuation based on the following criteria: (1) contained within the 3D lesion contour, (2) circular ROI with a diameter of 0.5 cm, and (3) maximal attenuation ≤300 Hounsfield units (HU) to exclude regions of calcification. The CAD algorithm quantified peak lesion attenuation in each phase each by reviewing all possible circular ROIs within the lesion contour and choosing the one with the highest mean Hounsfield unit value (Fig. 2). The lesion contours and appropriateness of CAD ROI placement were confirmed by consensus by two genitourinary fellowship-trained abdominal radiologists with 19 (SSR) and 10 (MLD) years of experience (who were blinded to all patient information, including the histopathologic diagnosis).

Fig. 2
figure 2

A 67-year-old man with a clear cell renal cell carcinoma. A The mass is heterogeneously enhancing on the corticomedullary phase. B Using an in-house-developed software, the entire lesion was segmented in the axial plane. C The CAD algorithm searched the volumetric lesion contour for the region of peak lesion attenuation contained within a circular ROI with a 0.5 cm diameter (red ROI)

Full size image

Peak lesion attenuation: manual ROI method

To establish the peak lesion attenuation values that were the reference standard for this study, one circular ROI of approximately 0.5 cm diameter was manually placed on the most avidly enhancing portion of heterogeneously enhancing lesions or in the center of homogeneously enhancing lesions in each post-contrast imaging phase. In addition, one circular ROI with a diameter of 0.5 cm was manually placed in adjacent uninvolved renal cortex in each phase. The lesion ROI and cortex ROI placement was confirmed by two genitourinary fellowship-trained abdominal radiologists by consensus.

Relative lesion attenuation

To normalize for the iodine load and variations in attenuation due to individual patient and technical factors, relative lesion attenuation was obtained by normalizing the CAD peak lesion attenuation and the reference standard, manual ROI peak lesion attenuation, to uninvolved renal cortex with the formula [(peak lesion attenuation ROI − cortex ROI)/cortex ROI] × 100%.

Statistical analysis

To compare the magnitude of CAD relative lesion attenuation of ccRCCs with the other renal masses in each of the four phases, we performed t tests. P values less than 0.01 indicated statistical significance and was adjusted for multiple comparisons (Bonferroni correction). A receiver operating characteristic curve (ROC) was used to evaluate the diagnostic performance of the CAD relative attenuation measurements. The area under the curve (AUC) was used to determine the threshold with the highest accuracy in discriminating ccRCC from chRCC, pRCC, oncocytoma, and fat-poor AML. For each threshold level, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy were calculated.

Bland–Altman analysis was performed to evaluate the level of agreement between the reference measurement (manual ROI lesion attenuation) and CAD lesion attenuation. The mean difference was used for measurement of the range of differences, expressed as 95% limits of agreement between the two methods. A receiver operating characteristic curve (ROC) was used to compare the diagnostic performance of the CAD relative attenuation measurements to the manual ROI method.

All analyses were performed by using software (IBM Corp. Released 2011. IBM SPSS Statistics for Windows, Version 20.0. Armonk, NY: IBM Corp.).

Results

Patients and lesions

A summary of the clinical and pathological characteristics for all patients and each lesion subtype are summarized in Table 1. Details of the study schema are illustrated in Fig. 1. Of the 200 patients in our study cohort, 130 (65%) were male, and 70 (35%) were female. The mean age at surgery or biopsy for all patients was 63 years (range 22–91 years). Of the 200 lesions analyzed, there were 106 (53%) ccRCC, 32 (16%) oncocytomas, 18 (9%) chRCCs, 34 (17%) pRCCs, and 10 (5%) fp-AMLs. For all patients, the mean time from preoperative CT to surgery or biopsy was 90 days (range 0–1061 days). Histopathology was diagnosed in 59 (29%) lesions from 20-gage percutaneous core biopsy, 103 (51%) from partial nephrectomy, 29 (15%) from total nephrectomy, and 9 (5%) from radical nephrectomy. The mean size for all lesions was 3.1 cm (range 0.8–6.9).

Table 1 Demographics and clinical characteristics of the study population
Full size table

Diagnostic utility of CAD relative lesion attenuation in discriminating clear cell RCC from the Other lesion subtypes

The CAD relative attenuation thresholds with the highest accuracy in discriminating ccRCC from each lesion subtypes in each phase are summarized in Table 3, as well as the AUC, sensitivity, specificity, PPV, NPV. The corticomedullary phase showed the greatest utility in CAD-derived relative lesion attenuation discriminating ccRCC from the indolent RCC subtypes and benign mimics. In this phase, ccRCC had a statistically significantly higher mean CAD relative lesion attenuation with no overlap in the 95% CI (20.3%, 95% CI 12.6–28.0) compared to chRCC (−20.5%, 95% CI −42.2 to 1.1, P = 0.002), pRCC (−41.3%, 95% CI −46.8 to −35.8, P < 0.001), oncocytoma (−11.7%, 95% CI −16.2 to −7.3, P < 0.001), and fp-AML (−16.8%, 95% CI −29.6 to −3.9, P < 0.001) (Table 2; Fig. 3) and an AUC ranging from 0.792 to 0.959 (Table 3). In the nephrographic and excretory phases, CAD relative attenuation also had a statistically significantly higher (P < 0.01) mean CAD relative lesion attenuation with no overlap in the 95% CIs compared to chRCC, pRCC, and fp-AML (Table 2; Fig. 3); however, the relative attenuation profile of ccRCC and oncocytoma had considerable overlap in these phases and was discriminated with accuracies of 42% and 55%, respectively (Fig. 3; Table 3).

Table 2 Mean CAD-derived relative lesion enhancement values for each histologic subtype of renal mass scanned on a multidetector CT using a four-phase renal mass protocol
Full size table
Fig. 3
figure 3

Bar graph represents the mean CAD relative lesion attenuation values on multiphasic multidetector CT for clear cell RCC, chromophobe RCC, papillary RCC, oncocytoma, and fat-poor AML. Error bars represent 95% CIs. CAD computer-aided diagnosis, CT computed tomography, RCC renal cell carcinoma, AML angiomyolipoma. * P value <0.01 compared to clear cell RCC relative attenuation

Full size image
Table 3 Receiver operating characteristic (ROC) results that represent the diagnostic performance of CAD relative lesion attenuation and the reference standard, manual ROI relative lesion attenuation, to discriminate clear cell RCC from the other histological subtypes
Full size table

Agreement between manual ROI and CAD ROI lesion attenuation measurements

Bland–Altman plots for assessing agreement between manual ROI peak lesion attenuation and CAD peak lesion attenuation methods are shown in Figure 4. No notable differences were observed between the two measurements, other than a slight positive bias in the CAD attenuation measurement. The mean difference between the two measurements was 14.0 HU in the corticomedullary phase, 25.5 HU in the nephrographic phase, and 26.6 HU in the excretory phase. ROC analysis of the diagnostic utility of manual relative lesion attenuation is shown in Table 3. Discrimination of ccRCC using CAD-derived relative lesion attenuation was comparable to the manual method for each lesion subtype in each phase with equal or higher AUCs and overlap in the 95% CIs of the AUC.

Fig. 4
figure 4

Bland–Altman plots for agreement between manual ROI peak lesion attenuation. CAD-derived peak lesion attenuation measurements had a mean difference of 14 HU in the corticomedullary phase (A), 25.5 HU in the nephrographic phase (B), and 26.6 HU in the excretory phase (C). ROI region of interest, CAD computer-aided diagnosis, HU Hounsfield units

Full size image

Discussion

The annual 2% increase in incidentally detected small renal masses (stage T1a) remains clinically challenging since all non-fatty solid enhancing masses are considered malignant until proven otherwise. Up to 20% of these T1a lesions are benign at surgical resection, and another 20% are less aggressive non-clear cell variants of RCC [1–7]. Unlike almost all liver, lung, and brain lesions, which are characterized routinely on imaging, histopathology on biopsy or nephrectomy remains the only accepted method for renal mass characterization [11, 13–15]. However, diagnosis by nephrectomy has several disadvantages. In addition to being costly, there are inherent intra-operative risks associated with anesthesia and surgery especially for patients with less aggressive RCC subtypes and benign tumors. There is increasing evidence that both simple and partial nephrectomy can worsen preexisting chronic kidney disease and increase cardiovascular comorbidities, especially in these often elderly patients [29, 30].

Renal mass biopsy has been an important diagnostic advance for characterization since the 1990s. Although relatively safe, biopsy has risk of bleeding and requires radiological and pathological subspecialty expertise for performance and interpretation. Characterization of renal masses from percutaneous biopsy also has variable accuracy ranging from 70% to 90% depending on tumor size, location, and physician experience and sampling technique, with a reported non-diagnostic biopsy rate between 10% and 30% depending on operator and pathological expertise [31, 32]. Retrospective studies have shown that if accurate and reliable pre-treatment knowledge of tumor histology and phenotype were available, the results would lead to an alteration from therapy (including surgery or tumor ablation) to active surveillance in up to 60% of cases [33–37]. Because 70% of renal lesions are detected incidentally on imaging, therapy is a clinical need for a robust method to obtain objective and quantitative data to assist in characterizing renal neoplasm histology on imaging.

In this study, we showed that a novel, quantitative in-house-developed CAD algorithm reliably detected peak lesion HU in a standard ROI on multiphasic MDCT and mimicked the standard manual method of ROI placement for detection of peak lesion HU in turn enabling robust renal lesion characterization by providing multiphasic objective measurement of relative lesion enhancement in comparison to normal enhancing renal cortex. Previous studies using manual ROI placement have validated the correlation between attenuation on CT and histology, and the value of a uniform four-phase renal CT protocol by showing each renal mass has a distinct attenuation profile across the corticomedullary, nephrographic, and excretory phases. However, the variability in the lesion classification results and lack of consensus as to which phase can best distinguish ccRCC from the other lesions is likely due to small sample sizes and because their quantitative data were obtained with subjective manual placement of a ROI of variable sizes in the most enhancing portion of the lesion [19, 22, 25–28].

Additionally, our results show that when the automated CAD algorithm was used to quantify maximal lesion attenuation relative to the cortex, ccRCC was the only lesion to have positive value in the corticomedullary and excretory phases (Fig. 3; Table 2), and we were able to achieve greater differentiation with less spread in the 95% CIs than previous studies [19, 22, 25–28] which utilized subjective manual placement of a ROI of varying sizes to obtain their quantitative data. We also confirmed our results in our dataset by comparing the performance of our CAD algorithm in detecting the highest region of peak lesion attenuation to the reference standard, subjective manual placement by a radiologist in the most enhancing portion of the lesion. Bland–Altman analysis showed excellent agreement between the CAD and manual ROI peak HU measurements, with slight positive bias of the CAD measurements in each phase (Fig. 4). Because the CAD algorithm was designed to find the region of peak attenuation within the lesion contour, this confirmed our hypothesis that CAD detected the ROI with the highest attenuation likely not readily visually discernable for manual placement. Additionally, prior work has shown that a small ROI with a cross-sectional area of 0.5 cm2 has the highest accuracy in discriminating malignant renal lesions on multiphasic CT [38]; therefore, we were able to automate placement of a ROI of this size in an attempt to reduce the inherent inter- and intra-reader variation associated with repeated attempts to manually place a ROI of the same size on CT images.

Our CAD method showed the highest utility with regard to differentiation of ccRCC from its benign mimics, oncocytoma, and fp-AML, where there was no overlap in the 95% CIs of the mean relative attenuation in the corticomedullary phase. (Table 2; Fig. 3). The ROC analysis showed CAD-derived relative attenuation in the corticomedullary phase discriminated ccRCC from oncocytoma with an AUC of 0.792 (vs. 0.725 using manual method) and fp-AML with an AUC of 0.825 (vs. 0.621 manual method) (Table 3). The nephrographic and excretory phases also showed a high utility in discriminating the RCC subtypes and the benign mimics, particularly fp-AML with an AUCs of 0.886 (vs. 0.742 manual method) in the nephrographic phase and 0.876 (vs. 0.629 manual method) in the excretory phase. To our knowledge, this CAD algorithm used on multiphasic CT is the most robust non-invasive and quantitative method reported to date for discriminating ccRCC from malignant RCC subtypes and benign mimics

Several CAD algorithms on imaging have been integrated into clinical practice and assist radiologists with diagnosing and characterizing lesions in the lung, breast, liver, brain, and bowel; however to our knowledge, none has been published for renal mass characterization based on differential enhancement [39, 40]. Previous studies evaluated the utility of CAD algorithms for renal mass detection on imaging based on morphology and texture features, but did not characterize individual histological phenotypes or stratify patients with malignancies [41–45]. Other limitations of these studies include CT scans performed without a uniform four-phase renal protocol and lack of pathologically proven lesion histology. Additional studies [46, 47] utilized a CAD algorithm to classify renal tumors as either RCC or benign; however, classifying all RCC lesions together is of limited clinical value since knowledge of lesion subtype is critical for prognosis. In addition, the inherently low attenuation of pRCC skews the mean attenuation value when all RCC lesions are grouped together as ccRCC and chRCC have greater peak lesion attenuation profiles on MDCT, making their classification schema less clinically applicable.

Other studies have shown promise in developing a diagnostic biomarker on imaging [48–50]; however, these techniques require definitive knowledge that the patient has a renal mass, and require an additional CT examination. This is not only an added expense, but often requires specialized expertise and equipment, subjecting the patient to additional radiation, risk of injection, and exposure to additional contrast material. This may not be feasible in elderly patients, patients with renal dysfunction or other comorbidities. However, if prospectively validated, segmentation of the lesion and our CAD algorithm can be combined with texture features and morphological analysis to construct a predictive model using a multiparametric approach to classify lesions [51] when the initial diagnostic CT is obtained and the renal mass is first detected.

Our study had some potential limitations. First, due to the retrospective nature of the study, there was variability in CT equipment and timing of multiple phases over a 16-year data acquisition period. Second, the segmentation of renal lesions for this study required approximately 5 min of post-processing time, but is comparable to the time necessary for a radiologist to load the DICOM images and review each scan to determine the most enhancing portion of the lesion in each phase. In addition, segmentation of the lesion provides an ROI of the entire lesion and is the first step toward the development of a quantitative multiparametric approach of renal lesion classification. Third, we had to manual place the ROI in the cortex, but we hope to automate these features in future development. Fourth, the low incidence of chRCC and fp-AML in the general population precluded a large number of these lesion types being including in our analysis and thus may be a source of selection bias.

In conclusion, we have shown that a novel CAD algorithm enabled objective and robust quantification of peak renal lesion attenuation relative to uninvolved enhancing renal parenchyma enabling robust non-invasive discrimination of ccRCC from other malignant and benign solid renal masses mimicking results from manual methods. Given our findings, this CAD algorithm may be a preliminary step in the development of a multiparametric decision model which may serve as an adjunct in clinical decision making to guide appropriate management and triage patients with incidentally detected T1 renal lesions to active surveillance, preoperative biopsy, or ablative and surgical therapies. However, our study is in the proof of concept phase and our preliminary CAD results should be validated in a prospective study and confirmed in a larger and independent cohort.