Exploring Multiphoton Microscopy as a Novel Tool to Differentiate Chromophobe Renal Cell Carcinoma From Oncocytoma in Fixed Tissue Sections

The surgical pathology archives from our institution were searched for the years 2012–2015 for cases with the diagnosis of renal oncocytoma and chRCC. A total of 27 cases—that is, oncocytoma (n = 12), conventional chRCC (n = 12), eosinophilic variant of chRCC (ie, chRCC with characteristic nuclear features of conventional chromophobe but predominantly eosinophilic cytoplasm; n = 1), and atypical oncocytic renal neoplasm (n = 2)–were selected. Atypical oncocytic renal neoplasm is the term used here to describe renal tumors that have features that are on the borderline between oncocytoma and chromophobe renal cell carcinoma. The H&E-stained slides from all 27 cases were retrieved from surgical pathology. These slides were reviewed and marked for tumor by the study pathologist, and unstained, formalin-fixed, paraffin-embedded sections were ordered from their corresponding blocks. The unstained formalin-fixed, paraffin-embedded slides were deparaffinized just before imaging with MPM. The tumor areas to be imaged on deparaffinized slides were marked by matching them with their corresponding H&E-stained slides. This study was conducted following Institutional Review Board approval.

A commercially available Olympus FV1000MPE (Olympus America, Center Valley, Pennsylvania) multiphoton laser scanning microscope (MPLSM or MPM) was used for this study. The deparaffinized unstained slides, with a coverslip on the tissue section, were placed on the MPM stage. A drop of normal saline was applied on the slide for the water-immersion objective. The tissue sections were then excited at a 780-nm wavelength using a Ti-Sapphire mode-locked femtosecond pulsed laser (Mai Tai DeepSee, Spectra-Physics, Newport Corp, Irvine, California). For these studies, emission signals—that is, short-wavelength autofluorescence (SWAF), long-wavelength autofluorescence (LWAF), and second harmonic generation (corresponding to 3 different wavelength ranges as defined in Table 1)—were collected as 12-bit greyscale images using 3 separate multialkali photomultiplier tubes, and then color-coded to improve visualization, as shown in Table 1. In this manuscript, we use reduced nicotinamide adenine dinucleotide (NADH) to denote the fluorescence from both NADH and reduced nicotinamide adenine dinucleotide phosphate (NADPH), because their emission spectra are identical.²⁰ 

Because the images were also used for quantitative analyses in addition to visual interpretation, all samples were imaged under identical conditions (eg, photomultiplier tube voltage and gain settings, laser power, etc). Additionally, the dark current was recorded in all channels with the shutters closed, which was later subtracted from the images. All imaging for this study was conducted using an Olympus ×25/numerical aperture 1.05 water emersion objective, which is optimized for light throughput from ultraviolet to near-infrared wavelengths. This magnification provided a field of view equivalent to approximately 0.5 mm × 0.5 mm (800 pixels × 800 pixels) with a lateral resolution of 0.636 μm per pixel. The pixel dwell time was set to 4 μs per pixel. Further, 2× optical zooms of 25× images were acquired to better visualize intracytoplasmic details without sacrificing lateral resolution. Image stacks were acquired by translating the focal plane in the axial direction with a step size of 1 μm. Because histomorphologic heterogeneity exists within a given tumor, multiple areas (3–4) within the tumor (outlined by the study pathologist using the corresponding H&E-stained slide as a reference) were acquired.

A blinded analysis was performed by 2 attending uropathologists at our institution. One of the uropathologists had previous experience with reading MPM images from human tissues, including the kidney, whereas the other uropathologist was a novice reader. For the blinded analysis, the study pathologist selected 4 images each from oncocytoma and conventional chRCC (total = 8 images) to familiarize the uropathologists with the unique and distinct signatures of these tumors on MPM. These signatures were identified by the study pathologist by comparing MPM images with their corresponding H&E slides. Histomorphologic features (identifiable on H&E) such as cell size, nuclear shape, type and amount of cytoplasm, and perinuclear halo—as well as distinct MPM signatures (not seen on H&E) such as intracytoplasmic granulesand nuclear autofluorescence—were used for differentiating the tumor types. After this brief training, the pathologists were shown 3 test images from each of the 27 cases (total = 81 images) and asked to classify them as oncocytoma, conventional chRCC, or eosinophilic variant/atypical tumors. Both the training and test sets were shown independently to the 2 attending uropathologists. The training images were not included in the test set.

As described in the Results, visually distinguishable intracytoplasmic granules were observed in oncocytoma and conventional chRCC. We thus used morphometric analysis to further quantify the differences between these granules, in an attempt to develop an algorithm to differentiate between these tumors with less investigator bias.

The steps for image processing and morphometric analysis methodology are detailed in the supplemental digital content (see supplemental digital content containing 1 figure and 2 tables at www.archivesofpathology.org in the March 2018 table of contents) and summarized in a flowchart (Supplemental Figure 1). In summary, binary masks were generated for the intracytoplasmic granules and cytoplasm in the “maximum projection” image of each raw image, and 7 morphometric features—including rescaled mean of granule intensities in SWAF (RMGIS), rescaled mean of granule intensities in LWAF (RMGIL), rescaled SD of granule intensities in SWAF (RSDGIS), rescaled SD of granule intensities in LWAF (RSDGIL), rescaled mean of cytoplasm intensities in LWAF (RMCIL), rescaled SD of cytoplasm intensities in SWAF (RSDCIS), and rescaled SD of cytoplasm intensities in LWAF (RSDCIL), as detailed in Supplemental Table 1—were calculated for the mask image and used for subsequent analysis. Single features as well as combinations of features (feature vectors) were then used to distinguish conventional chRCC and oncocytoma by using support vector machine (SVM)²¹  and leave-one-out cross validation (LOOCV).^22,23  A set of statistical measures, including sensitivity, specificity, accuracy, and area under the curve of receiver operating characteristic (ROC) curve (AUROC), were used to evaluate the classification performance,^24,25  as detailed in Supplemental Table 2. A nested cross-validation,^26,27  which included an internal and external LOOCV, was used to find the optimal feature combination for classification and its performance. The optimal feature vector V_optimal, which provided the best performance in distinguishing the 2 tissue types, was determined using the internal LOOCV. The accuracy of the classification for V_optimal provided by the external LOOCV was used as the final diagnostic accuracy for morphometric analysis.

Figure 1, A and B, shows the various MPM signatures of chRCC. These include: (1) predominant sheetlike or trabecular growth pattern, and (2) large cells with distinct cell borders, wispy scant cytoplasm (SWAF; color-coded green), and pleomorphic nuclei. The nuclei had a variable amount of autofluorescence in SWAF channel (color-coded green), similar to the cytoplasmic signal. In some cases, perinuclear halo was also identified as a signal-void area around the autofluorescent nucleus. These MPM features resembled morphologic features seen on the corresponding H&E (Figure 1, C and D). However, in addition to the above-mentioned morphologic features, the most prominent and unique feature was the presence of intracytoplasmic granules in these tumors, not seen on H&E. The granules were small, “sandlike” in appearance, with autofluorescence in both SWAF and LWAF channels (appearing bluish green in our color code). These granules did not have any specific cytoplasmic localization but were seen diffusely scattered in the cell cytoplasm. Based on the distribution of these granules and prior electron microscopy studies,^8,9  we hypothesize that this signal originates from the microvesicles known to exist in chRCC cytoplasm. In addition to the cellular autofluorescence, second harmonic generation signal (color-coded red) was also identified originating from the stroma, comprising primarily collagen fibers.

Figure 2, A and B, shows the various MPM signatures of oncocytoma. Oncocytomas had a nested or acinar pattern, with a uniform population of small cells. These cells had abundant, homogenous cytoplasm (SWAF; color-coded green), with a central signal-void nucleus. Again, these MPM features resembled morphologic features seen on the corresponding H&E (Figure 2, C and D). Similar to conventional chRCC, intracytoplasmic granules were also identified in oncocytoma by MPM but not H&E. However, unlike granules of conventional chRCC, the granules of oncocytoma appeared larger and brighter, with autofluorescence especially in the LWAF channel (color-coded blue). In contrast to conventional chRCC granules, the oncocytoma granules had a specific cytoplasmic distribution pattern (ie, apical, and/or perinuclear). Based on the distribution of these granules, we hypothesize their origin from mitochondria in oncocytoma, as has been previously documented on electron microscopy.⁹ 

Similar to conventional chRCC, collagenous tissue (second harmonic generation; color-coded red) was also identified in oncocytoma stroma. In our study, no significant difference was seen in the distribution pattern or other morphologic features of the collagen in these 2 tumor types. Table 2 shows a comparative analysis of chRCC and conventional oncocytoma as seen on MPM and corresponding H&E.

In addition to the above-mentioned typical cases of conventional chRCC and oncocytoma, we also imaged 3 variants of oncocytic tumors—that is, eosinophilic variant of chRCC (n = 1) and atypical oncocytic renal neoplasm (n = 2)—as diagnosed on H&E. Representative images from these tumors are presented in Figure 3. The eosinophilic variant of chRCC (Figure 3, A and B) had morphologic features similar to those of typical oncocytoma (ie, uniform population of cells with abundant cytoplasm). However, unlike oncocytoma, autofluorescent nucleus and perinuclear halo (signal-void area surrounding autofluorescent nucleus) were seen in this tumor. The granules of the eosinophilic variant of chRCC had characteristics of both tumor types (Figure 3, A and B). On the other hand, the atypical oncocytic renal neoplasm (Figure 3, C) had MPM signatures resembling those of typical oncocytoma (as mentioned above) and was difficult to differentiate from the latter. These MPM features resembled morphologic features seen on the corresponding H&E (Figure 3, D through F).

Based on the unique MPM signatures and above-mentioned histomorphologic features, both of the uropathologists could classify all cases of oncocytoma (12 of 12; 100%) and conventional chRCC (12 of 12; 100%) accurately. The uropathologist with experience in MPM was able to subclassify 1 of the chromophobe RCCs as an eosinophilic variant, whereas the novice uropathologist could not detect a particular subclassification of this lesion. On the contrary, because of the very close resemblance of histomorphologic features of oncocytoma with atypical oncocytic renal neoplasm, these tumors were classified as oncocytoma (0 of 2) by both of the readers. Nonetheless, this finding suggests that MPM could be an additional diagnostic modality to prognosticate these indeterminate lesions. Studies on larger scales are warranted to affirm this hypothesis, along with the follow-up of the cases for recurrence. An overall diagnostic accuracy of 100% was achieved by both of the uropathologists in differentiating conventional chRCC from oncocytomas. Because of the small sample size of the variant cases, they were not included in the diagnostic accuracy evaluation.

Typical maximum projections of stack images for conventional chRCC and oncocytoma specimens are shown in Figure 4, A and D, respectively. The binary masks of granules and cytoplasm are shown in Figure 4, B and C, for conventional chRCC, and Figure 4, E and F, for oncocytoma, respectively.

Using the nested cross-validation for case-based data set, the overall optimal feature selection was found to be V_optimal = (RMGIS, RSDCIS). In other words, a combination of the mean granule autofluorescence intensities in the short-wavelength channel and the SD of the cytoplasmic autofluorescence intensities in the short-wavelength channel, both rescaled by the mean cytoplasm intensity in the short-wavelength channel, best separated the conventional chRCC cases from the oncocytomas. The overall accuracy for internal LOOCV was calculated to be 83.0%, along with overall sensitivity of 74.2% and overall specificity of 91.7%. The sensitivity, specificity, and accuracy of the final evaluation of the classifier with V_optimal were calculated to be 75.0%, 91.7%, and 83.3%, respectively. The linear SVM classifier trained using the V_optimal of the whole data set is shown in Figure 5, A. The ROC curve for SVM classification using V_optimal is shown in Figure 5, B. The AUROC for SVM was found to be .85 (as a comparison, the AUROC value for a perfect separation between 2 classes is 1, whereas the value for a random guess separation is .5). All of the above statistical measures for feature selection and final evaluation of the classifier are shown in Table 3. For comparison, several other top selections of features were also considered. In particular, the second optimal selection was found to be V⁽²⁾_optimal = (RMGIS, RSDGIS, RSDCIS), which provided an overall internal LOOCV accuracy of 82.6%, and a predictive accuracy for external LOOCV of 83.3%. The corresponding AUROC was also calculated and shown in Table 3.

Based on the SVM classifier, which was trained with the V_optimal of all 24 cases of conventional chRCC and oncocytoma, the cases of eosinophilic variant of chRCC and atypical oncocytic renal neoplasm were then classified into these 2 tumor types (chRCC and oncocytoma). The only case of eosinophilic variant of chRCC was classified as oncocytoma, and the 2 cases of atypical oncocytic renal neoplasm were split with diagnosis as chRCC and oncocytoma by morphometric analysis.

In summary, the results showed that typical cases of chRCC and oncocytoma can be distinguished with 83.3% accuracy by automated morphometric analysis. We did not include the 3 variants of oncocytic neoplasms in this diagnostic accuracy evaluation because of the small sample size. By combining multiple parameters (features) of the samples, better discrimination was achieved compared with using individual parameters. Current results show that using the 2 features (RMGIS, RSDCIS) seems to be adequate to provide the best separation between the 2 types of tissues. The accuracy of morphometric analysis is somewhat lower than the blinded pathologist's analysis. We attribute this to the different parameters used in these 2 methods, that is, morphometric analysis only used intensities of granules and cytoplasm, whereas blinded pathologist's analysis used a constellation of morphologic features. Quantification of other morphologic features will be performed in our future study.

Oncocytic tumors are at times difficult to diagnose on H&E and may require ancillary studies, such as immunohistochemistry, molecular analysis, and electron microscopy. However, despite a huge arsenal of ancillary tools studied to date, the distinction between various oncocytic tumors remains challenging and time-consuming. Thus, novel optical imaging tools, such as MPM, that image tissue at cellular resolution are worth pursuing as alternative diagnostic tools for oncocytic tumors.

Multiphoton microscopy has been used in research to differentiate neoplastic from nonneoplastic tissue and to characterize various tumors.^15–18  Further, Jain et al¹⁹  have characterized various malignant kidney tumors on MPM and have achieved a high diagnostic accuracy of 98% in subtyping these tumors. Although this study included few cases of chRCC, it lacked oncocytoma cases. Thus, our current study builds on this previous work to address the critical need for rapidly and accurately differentiating chRCC from oncocytoma.

The purpose of this study was to explore the potential of MPM as a rapid diagnostic tool in differentiating oncocytoma from chRCC. Because this was a “proof-of-concept” study, mostly typical cases (those readily diagnosed on H&E) of oncocytoma and chRCC were imaged to characterize their MPM signatures. Based on the histomorphologic features and unique MPM signatures, we could reliably differentiate oncocytoma from conventional chRCC. The histomorphologic features of typical oncocytomas and conventional chRCCs resembled those seen on H&E slides. However, the most striking MPM signature of both of the tumors was the presence of autofluorescent intracytoplasmic granules, which are not seen on H&E-stained slides. Although we saw these granules in both types of tumors, they appeared to be distinct from each other based on their size, shape, cytoplasmic distribution, and autofluorescence wavelengths, and thus were valuable in arriving at a definitive diagnosis. Based on intracytoplasmic distribution of these granules and previous electron microscopy studies,^7–9  we hypothesize their origin from microvesicles in chRCC and mitochondria in oncocytoma. Generally, oncocytoma is rich in mitochondria; the reason that they appeared sparse and large on MPM could be that this technology is detecting only certain  abnormal mitochondria. Although the eosinophilic variant of chRCC had many intermediate features, it could be correctly classified by the experienced MPM reader, particularly through evaluation of the granules on MPM. Both atypical oncocytic renal neoplasms were classified as oncocytoma on MPM, suggesting the possibility of using MPM in arriving at a definitive diagnosis (benign versus malignant) for intermediate tumors on H&E. The fact that a high diagnostic accuracy and 100% concordance was achieved by both the experienced and the novice MPM reader demonstrates the reproducibility and the reliability of the MPM signatures defined in our study. To test the possible automation of MPM diagnosis, a morphometric analysis was performed that also yielded a high diagnostic accuracy of 83.3%. The sensitivity and specificity were found to be 75% and 91.7%, respectively. Sensitivity was a little lower than specificity, which may be attributed to higher variation in the features of chRCC samples than those of oncocytoma. Furthermore, atypical oncocytic renal neoplasms had variable values on morphometric analysis: one was classified as oncocytoma and the other as chRCC. This variation could be the result of small sample size in our study. Another method to evaluate the performance of the classifier is AUROC. Its value of .85 means the probability that the classifier trained from the current cases ranks a chRCC (1) sample higher than an oncocytoma (0) sample is .85, that is, the probability of correctly distinguishing chRCC and oncotycoma is 85%.

To the best of our knowledge, our group has for the first time demonstrated the potential of MPM in differentiating oncocytoma from chRCC. Stewart et al²⁸  have demonstrated the capability of Raman microscopy imaging to differentiate between oncocytoma and chRCC and demonstrated a sensitivity and specificity of 86% and 81%, respectively, in their study. However, the major limitation of this technique is the long image acquisition time for a field of view compared with the acquisition time for MPM. For example, based on the parameters used by Stewart et al²⁸  in their study, it will take more than 6 hours for Raman microscopy imaging to acquire the proposed hyperspectral image stack with an approximately 0.5 × 0.5 mm field of view. In contrast, MPM can rapidly scan a similar field of view within a few seconds. Thus, Raman microscopy imaging may not be a suitable tool for rapid real-time diagnosis, particularly when multiple sites need to be imaged within in the same specimen to account for tumor heterogeneity, which has been especially noted in oncocytic tumors.²⁹  Furthermore, we obtained a much higher diagnostic accuracy on blinded analysis as well as on morphometric analysis compared with Raman microscopy imaging.

Although in our study we could reliably differentiate oncocytoma from conventional chRCC, it is not without limitations. One of the major limitations was the inclusion of mainly typical oncocytoma and chRCC cases (ie, readily diagnosed cases on H&E) with an exception of a few variants of oncocytic tumors. In real clinical practice, it is these variants that most often pose diagnostic dilemmas, requiring expensive and time-consuming ancillary evaluations. Another limitation was the small sample size of the tumors used in this study. Lastly, eosinophilic variants of other kidney tumors, such as clear cell RCC and papillary RCC, that are considered in the differential diagnosis of these tumors were not imaged in our study. We hope to remedy these limitations in our future studies.

Multiphoton microscopy is a potential diagnostic tool to differentiate between chRCC and oncocytoma solely based on their unique MPM signatures and may aid in the clinical management of the patient.