Abstract
Background: Response Evaluation Criteria in Solid Tumors (RECIST)-defined measurements are limited when evaluating soft tissue sarcoma (STS) response to therapy. Histopathological assessment of STS response requires a determination of necrosis following resection. A novel semi-automated technique for volumetric measurement of tumor necrosis, using enhanced magnetic resonance imaging (CE-MRI), is described. Patients and Methods: Eighteen patients with STS were treated with neoadjuvant therapy and then resected. CE-MRI, obtained prior to resection, were evaluated by two observers using semi-automated segmentation. Tumor volume and percent necrosis was compared with histology and RECIST measurements. Results: The median percent necrosis, determined histologically and from CE-MRI, was 71.9% and 67.8%, respectively. Accuracy of these semi-automated measurements was confirmed, being statistically similar to those obtained at histopathological assessment of the resected tumor. High Intra-class correlation co-efficients suggest good inter-observer reproducibility. Tumor necrosis did not correlate with the RECIST measurements. Conclusion: Semi-automated determination of tumor volume and necrosis, using CE-MRI, is suggested to be accurate and reproducible.
Soft tissue sarcoma (STS) comprises a diverse group of tumors of mesenchymal differentiation, which all together account for approximately 10,000 cases per year in the United States (1). Patients with some STS receive external-beam radiotherapy (RT) and/or chemotherapy prior to surgical resection with curative intent (2). At present, the response of the tumor to chemotherapy/radiation is not reliably identified until histological assessment of the resected tumor. Imaging-based, non-invasive, detection of meaningful tumor response to therapies has proven elusive (3, 4). Standard measurements of the single longest diameter of the tumors, defined by the Response Evaluation Criteria in Solid Tumors (RECIST), have been shown to correlate poorly with oncological outcomes in many tumor types (5, 6, 7). This is well-described in gastrointestinal stromal tumors (GISTs), being treated with imatinib (8). New criteria have been recently proposed for GISTs incorporating tumor attenuation in addition to tumor size (9, 10, 11). These criteria have been suggested to correlate with patient outcomes better than linear measurements of the whole tumor, as defined by the RECIST criteria. With regard to STS, it has been long-appreciated that there may be a dramatic histological response without any decrease in size, as measured in accordance with the RECIST criteria (12, 13, 14).
It has been suggested in a number of tumors that volumetric determination of tumor necrosis, on imaging studies, may better-predict survival (15, 16, 17). A quantitative non-invasive, estimate of STS tumor necrosis during various stages of chemotherapy or radiotherapy may yield prognostic information that is useful for subsequent patient treatment. Enhanced cross-sectional imaging, such as CE-MRI, allows for gross evaluation of the entire heterogeneous tumor volume, while histopathological determination of necrosis for STS may be based on selected sections of the tumor, chosen by the pathologist and may not accurately represent the entire tumor volume.
In the current study semi-automated methods for volumetric segmentation of viable, enhancing, and necrotic, non-enhancing, portions of tumor are applied to CE-MRI of STS. The goal of this pilot study is to assess the accuracy and inter-observer reproducibility of this technique.
Patients and Methods
Approval for this study was obtained from the Institutional Review Board. The study was conducted in compliance with Health Insurance Portability and Accountability Act (HIPAA) guidelines. Clinical, histopathological, imaging and treatment data were retrospectively reviewed. Between March 2008 and February 2011, a total of 18 patients with locally-advanced, non-metastatic STS of the extremity of varying histology (Table I), were treated with neoadjuvant therapy, then resected. These patients were prospectively entered and tracked in a computerized cancer center database. A diagnosis of soft tissue sarcoma was established by image-guided core biopsy. Patients with intermediate or high-grade tumors greater than 5 cm, patients with low-grade tumors > 8 cm, or patients with tumors abutting key neurovascular structures were treated with neoadjuvant radiation therapy (RT) monotherapy. RT was administered in 2-Gy fractions, over 25 sessions for a total dose of 50 Gy. Surgical resection was performed approximately 4 to 6 weeks following the completion of RT to allow acute effects of radiation, such as dermatitis, to resolve. Seven patients were treated with the multi-kinase inhibitor sorafenib in combination with neoadjuvant RT, on a phase I trial. All patients underwent limb-sparing resection with curative intent, and pathologic margins were tumor-negative. CE-MRI images, obtained prior to and following treatment, within three weeks of resection, were evaluated with semi-automated segmentation techniques to determine whole tumor volume and percentage of, non-enhancing, tumor necrosis. These measurements were compared to the percentage of histologically-determined necrosis of the surgically resected tumor, as well as the longest tumor diameter, determined in accordance with the RECIST criteria (18) (http://ctep.cancer.gov/protocolDevelopment/docs/recist_guideline.pdf). The Electronic Medical Record (EMR) and imaging studies stored on a Philips iSite Picture Archiving and Communication System (PACS, Philips Healthcare, Andover, MA) were used to retrospectively determine time-to-distant recurrence and disease-free survival rates.
CE-MRI. MRI was obtained at the time of diagnosis and as close to surgical resection as possible. Ideally repeat-imaging was performed 2 to 3 weeks before surgery to allow for resolution of post-radiation changes and to help distinguish therapeutic effects, which may evolve over time, from side-effects, such as tumoral edema, to resolve. However, in some cases this was not possible due to patient availability and MRI scheduling. The mean time interval between the MRI and surgery was 13.6 (SD: 11.3) days with a range of one to forty-six days. Four patients had an interval greater than three weeks, 22 to 46 days, and were not included in determinations of accuracy, comparing imaging to histology.
All MR imaging examinations were performed with a 1.5-T MR system (Signa Horizon EchoSpeed HDxt, LX Version 15 operating system software; GE Healthcare, Inc., Milwaukee, Wis). For each patient, a standardized imaging protocol was performed on the same scanner. This included fat-saturated T1-weighted MR imaging (repetition time msec/echo time msec of 400/17) and fat-saturated T2-weighted MR imaging (4,000/85) in the transverse, sagittal and coronal planes. The coronal plane was usually used for the dynamic enhanced portion of the study to view the feeding vessel all in one section. Initially, gadopentetate dimeglumine (Magnevist; Berlex Laboratories, Wayne, NJ) was administered via a peripheral arm vein, at a concentration of 0.1 mmol per kilogram of body weight, resulting in a standard dose of 1 mL of gadopentetate dimeglumine per 10 kg of patient weight. A 20-mL saline flush was administered at the same flow rate immediately after the gadopentetate dimeglumine. A power injector delivery system (Medrad, Indianola, Pa) was used to provide fixed flow rates of 2 mL/sec. The phase of imaging showing greatest conspicuity of enhanced and necrotic, non-enhancing tumor, were chosen for analysis. MRI images were stored on a Philips iSite Picture Archiving and Communication System (PACS, Philips Healthcare, Andover, MA) and were then de-identified in accordance with the HIPPA requirements. The observers were blinded to patient demographics.
Semi-automated determination of whole-tumor volume and the volume of necrotic tumor. The two observers (B.J. and C.M.) first received training regarding the appearance of enhancing and non-enhancing sarcoma and the use of the software for semi-automated calculations of tumor volumes and the volume of non-enhancing, necrotic, tumor tissue. The training session was conducted by an attending radiologist (W.L.M.) with 10 years of experience and included analysis of 10 CE-MRIs of STS, following treatment and just prior to resection. The 10 MRI scans used for this training were not included in the study. To avoid bias during measurements there was no consensus reading and observers were blinded to histopathological data and to each others' measurements.
Our semi-automated segmentation software was developed on a Windows XP (Microsoft Corporation, Redmond, WA) workstation using the MATLAB programming language (version R2007a; The MathWorks, Natick, MA). The program was run on an HP Pavilion dv1000 Laptop with a 1.5-GHz processor, 2 GB RAM, and an 80-GB hard drive (Hewlett Packard, Palo Alto, CA). A graphical user interface was implemented. The semi-automated segmentation software, developed for volumetric analysis, used algorithms based on Iterative Watershed (IWS) (19) and Pixel Thresholding (PT) (20). The IWS algorithm implemented, first used gradient vector flow transformation of the image for edge enhancement, which was followed by a region-growing method, analogous to “flooding” the gradient vector flow map. Sarcoma tumor edges were finally determined, on each slice, from resultant formed ridgelines (Figure 1), as previously described and validated by Ray et al. (19), and then further applied to hepatic tumors undergoing radioembolization and chemoembolization (15, 16).
Briefly summarized; segmentation was performed by placing a series of markers that defined the outside of the tumor as well as series of markers that defined a representative sample of the tumor parenchyma. The software automatically connected each marker with a straight line, creating a polygonal shape (Figure 1a). On each axial image, the operators had the option of checking their work by way of an output image or to continue to the next set of images. The operators had the ability to delete, add, or move any marker at any time. When all image markers were set for all individual slices, the tumor edges were automatically outlined (Figure 1b), as described by Ray et al. (19). The program added the individual slice areas and total volume were recorded accounting for the voxel dimensions.
The IWS algorithm calculated the whole-tumor volume (enhancing and non-enhancing). This was followed by application of the Pixel Thresholding (PT) algorithm, in which a region of normal parenchyma and the enhancing portion or necrotic non-enhancing portion of the tumor were included in separate regions of interest. The histogram of pixel attenuation values was evaluated (Figure 2a). An automatic technique of analysing histograms similar to Otsu's method is initially applied and the observer could then select a pixel attenuation value as the cut-off were all pixel values >x are counted as enhancing viable tumor which resulted in the image most clearly distinguishing enhancing viable tumor from non-enhancing necrosis based on the resultant outline and the histogram (Figure 2b).
Histopathology. Hematoxylin and Eosin–stained slides were manually reviewed in a blinded fashion by a pathologist (D.B.). In all except two tumors the gross specimen was oriented in the coronal direction, and cut at the midline, and center, of the tumor. As part of the usual examination of these tumors, a 5-μm thick slice parallel to the first cut was removed for histological analysis. The percentage of histologically-intact tumor and the percentage of necrotic tumor were scored per slide, and the mean percentage of tumor necrosis was calculated, using a continuous scale. The median number of stained slides examined per patient was 10 (range, 8–18). A clinically good response is considered to be at least 90% necrosis (i.e., grade III or IV) (3).
In two tumors, a more complex and complete sectioning of the tumor was performed. The spatial coordinates from which slides were prepared for histological analysis were defined and plotted on a map. The coordinates and maps were then used during analysis of imaging studies to determine the percentage of necrosis at each of these regions of the tumor (Figures 3 and 4).
Statistical analysis.
Interobserver reproducibility of measurements of the volume of necrosis. The intra-class correlation co-efficient (ICC) (21), for the two observers was calculated to determine inter-observer reproducibility for the measurements of the volume of sarcoma necrosis (%), determined using described semi-automated analysis of CE-MRI. Confidence limits were set at 95%. We also constructed Bland–Altman plots (22) between observers, in which the graphs were assigned the mean of the volumes, as measured by the two operators, as the abscissa (x-axis) value and the difference between the two necrosis volumes as the ordinate (y-axis) value. The Spearman's rank correlation co-efficient was used to study the correlation between tumor necrosis (%), determined at histopathology and imaging, RECIST-defined change in longest tumor diameter (%), and percent change in whole-tumor volume, determined from segmentation of CE-MRI
Results
Inter-observer reproducibility. Initial validation of this approach requires a measurement of inter-observer reproducibility. Semi-automated measurements of the percentage of tumor necrosis are made by 2 observers and compared. For 18 tumors analyzed, observer one and two calculated a median necrosis percentage of 63.7 (SD: 25.7) and 67.8 (SD:22.3), respectively. An ICC of 0.93 and 95% confidence limit of 0.83 and 0.97, respectively, suggest a high degree of inter-observer reproducibility. The Bland and Altman plot also suggests good inter-observer reproducibility (Figure 5).
Accuracy of semi-automated segmentation to determine percentage of tumor necrosis and whole-tumor volume. Non-enhancing portions of the tumor are thought to be non-perfused and likely non-viable and necrotic. In an attempt to initially confirm accuracy of this determination, comparisons are made to the histopathological assessment of resected tumor necrosis, serving as the gold standard. For 18 tumors analyzed, observer one and two calculated a whole tumor median necrosis percentage of 63.7% (SD: 25.7) and 67.8% (SD:22.3), respectively, with a range of 21 to 96% and 25 to 96%, respectively. Histological assessment demonstrated a median percent necrosis of 71.9% (SD: 31.4) with a range of 5 to 100%. Seven tumors demonstrated near-complete pathological response (≥95% necrosis) (36%).
There where four patients that had greater than a three-week interval between MRI and resection, as a result of patient non-compliance with scheduling of imaging and surgery. If these patients are not included when determining accuracy, as defined as an agreement between the imaging and histology results, a high degree of accuracy is suggested. The ICC is 0.92 with 95% confidence limits of 0.78 and 0.97, respectively. The Bland and Altman plot comparing percentage of tumor necrosis, determined by segmentation of CE-MRI with necrosis determined histologically also suggests excellent agreement of the measurements (Figure 6).
Usual histopathologic determination of tumor necrosis involves sampling of the mid-portion of the tumor and hence there may be propensity for sampling errors. One potential advantage in determining the percentage of tumor necrosis, based on enhanced cross-sectional imaging, is that the entire tumor volume is evaluated potentially eliminating sampling error. However, we found that measurements of tumor necrosis made on three consecutive MRI slices, obtained at the center/midline of the tumor (mean±SD: 66.0±26.8), were not significantly different from the percentage of necrosis obtained with analysis of the whole tumor volume (mean±SD: 63.7±25.7). When comparing MRI measurements at the midline of the tumor to histological measurements the ICC of 0.91, 95% confidence limit of 0.76 and 0.97, and the Bland and Altman plot suggest good agreement.
For two tumors, a more laborious complete and detailed approach to histological evaluation is utilized including detailed graphical depiction of locations within the tumor from which sections are taken for histological analysis (Figs 3 and 4). We sought to measure necrosis using segmentation of CE-MRI at specific locations were histological determination of necrosis was performed, using the grid representation of locations of tumor sectioning, with measured coordinates, to determine theses locations on CE-MRI images. For these two tumors 7 sections were obtained from which histologic sections were prepared and CE-MRI segmentation was performed. These measurements were very similar with MRI-determined necrosis mean measurements of 82±11% and histologically-determined mean measurements of 89±7% necrosis.
The accuracy of semi-automated determination of the volume of the whole tumor, is evaluated by comparing these measurements to the actual volume of the resected tumor. Three tumors were predominately cystic and a large amount of cyst fluid spilled from the specimen when being measured. Therefore, actual measurements of these tumors were deemed to be erroneous and are not included in this analysis. However, comparisons using the other fifteen tumors suggest a high degree of accuracy of the volume determinations. The mean tumor volume, calculated using the semi-automated segmentation was 501.2±571 and actual mean volume, determined at pathology, was 448±568.5. The ICC of 0.98, 95% confidence limits of 0.94 and 0.99, as well as the Bland and Altman Plot, all demonstrate excellent agreement between the tumor volume obtained from MRI segmentation and the actual tumor volume determined at pathology.
Comparison of RECIST criteria defined, linear measurements, with volumetric measurements of the whole tumor and tumor necrosis. In accordance with RECIST, the single longest linear diameter of each tumor is manually measured on our PACS system, by one of the observers (B.J.). Measurements were made from CE-MRI images obtained at the start of treatment and within three weeks of surgical resection. The whole tumor volume and percent change, over the course of treatment, is also determined from these MRI images, using semi-automated segmentation. The median tumor diameter was 10.3 cm (range, 2.2-18.7 cm). The median percent change in tumor diameter is −11.2 ±19.8% (range, −33.6 to +32.3%), while the median change in tumor volume is −10.3 ±60.5 cm3.
There were no statistically significant correlations between RECIST-defined measurements and necrotic tumor measurements, suggesting that RECIST criteria may not reflect clinically-relevant changes in tumor necrosis. However, there was a statistically significant correlation between tumor necrosis (%), and tumor volume (cc), both determined from CE-MRI, obtained within three weeks of resection (Spearman's rank correlation co-efficient=0.537; p-value=0.021). There was also a statistically significant correlation between the percent change in the longest tumor diameter and the percent change in whole-tumor volume (Spearman's rank correlation coefficient=0.847; p-value<0.0001).
Discussion
Currently, quantitation of necrosis in the surgical specimen is the most commonly applied assessment of response available to clinicians treating solid tumors with neoadjuvant therapy (23). Studies in STS, as in other tumors, may suggest degree of necrosis to correlate with survival (24). Although these studies have been retrospective in nature, have analyzed varying thresholds of tumor necrosis (80-99%), and have employed heterogenous pre-operative regimens including multi-agent chemotherapy, chemoradiation, or RT monotherapy, the degree of treatment-induced pathologic necrosis in the surgical specimen continues to be used as a short term quantitative end-point in STS, despite limited data establishing a link between extent of pathologic necrosis and long term outcomes (25). It is useful to have a reproducible and validated surrogate imaging end-point for early determination of response. The purpose of the present study is to describe a novel semi-automated segmentation approach applied to CE-MRI to determine STS volume and the percentage of necrosis, non-enhancing, tumor. This pilot study represents initial steps to validate this approach, determining its accuracy and reproducibility (26).
STS show significant histological treatment responses in the form of hyaline fibrosis, necrosis and granulation tissue. Despite this, there is minimal early change in the overall size of the tumor, as evaluated using RECIST criteria, and volumetric whole-tumor measurements (14, 27). Canter et al. have previously shown that these linear measurements did not predict clinical outcome while tumor necrosis, determined with histopathology, did predict clinical outcome (14). This suggested the need for novel imaging end-points to evaluate response without the need for resection. For example, the Choi criteria of evaluating tumor contrast-enhancement at CE-MR or CE-CT, are said to complement tumor size determination and have been suggested to be predictive of pathologic response (10).
The measurements described here are likely to predict clinical response since local treatment failure, evidenced by lack of substantial necrosis, frequently contributes to patient death (28). However, power calculations suggest that at least 96 patients would be required to establish the ability of this approach to predict survival, which is out of the scope of this initial description and validation. Given the small sample size of the current study, it is not certain that the results can be generalized for all soft tissue sarcomas. It is important to note that biological plausibility-alone is inadequate to allow any end-point to be validated without a demonstration of correlation with true patient benefit outcomes (26). This suggests the need for further evaluation of the technique in large prospective Phase II or III clinical trials following this initial validation, as proof-of-concept.
Assessment of necrosis, as determined from histology, utilizes 5 micron slices, usually from the midline of the tumor, which the pathologist feels is representative of the entire tumor. With this usual histological determination of tumor necrosis, it is likely that there is the potential for sample error, in which the chosen tumor sections may not be representative of the entire tumor necrosis. It is known, from enhanced imaging, that often times the pattern of necrosis is patchy and irregular in shape throughout the tumor volume (15, 16). We hypothesized that the imaging-based evaluation of the whole tumor's necrotic volume would reduce this sampling error. However, we found that the percentage of non-enhancing necrosis, throughout the whole tumor, was not significantly different from necrosis at the midline of the tumor, corresponding to the section of the resected tumor usually evaluated histologically. This suggested that the sampling error may be less of an issue than originally hypothesized.
A potential limitation of this approach is that imaging slices are approximately 5 to 10 mm and therefore do not directly match the histopathologic slices, perhaps limiting comparisons of imaging with histology. Furthermore, CE-MR imaging demonstrates lack of tissue enhancement, thought to correspond to necrosis, and not actual cell death, as demonstrated histologically. To further investigate the accuracy of the imaging-based evaluation of tumor necrosis we evaluated more sophisticated approaches of documenting tumor section locations, by mapping coordinates of where tumor sections were taken (Figures 3 and 4). Determination of “necrosis”, based on CE-MRI, also correlated well with histological determination at these specific sections. It is, thus, suggested that areas of non-enhancing, low-signal intensity tumor, likely correspond with tumor necrosis determined with H&E staining of tumor sections. While our results suggest good correlation of percentage of non-enhancing, necrotic, tumor on MRI with that determined at histology, Bajpai et al. found that a change in the pattern of enhancement did not correlate with necrosis (29).
A possible limitation of CE-MRI is that over time Gadolinium may diffuse into the necrotic tissue, thereby potentially obscuring tumor necrosis on the contrast-enhanced MR images. This may result in false reduction in “necrotic volume” (31). In the current study, there was likely little diffusion of the gadolinium across the interstitium into the necrotic tissue, since images are obtained in the arterial phase of enhancement, 4 sec following contrast administration. Another limitation of CE-MRI is that portions of a tumor may enhance and yet contain no viable tumor cells; for example, the supporting stroma may persist despite death of the tumor cells, and enhancing reparative granulation tissue may be present. This could result in falsely-elevated determination of “viable” tumor tissue. Our results suggest good accuracy of this approach for the depiction of necrotic tumor tissue, correlating-well with percentage tumor necrosis determined histologically.
Measurements of whole-tumor volume may be useful in evaluating response to therapy (18, 30). However, these approaches also need to be validated. We found that whole-tumor volumes determined with these semi-automated techniques corresponded well with tumor volumes determined by pathologists, following resection. Any discrepancy between volume change, as assessed by MRI, and at histopathology might be explained by the slow regression of any osteoid matrix and/or cystic degeneration in good responders with possible corresponding fallacious increase in size on MRI (32). This may be illustrated in the three resected cystic tumors that could not be accurately measured at pathology.
When evaluating reproducibility, two blinded medical students performed the analysis, following a training session. The intent was to demonstrate that relatively inexperienced observers could accurately perform this semi-automated segmentation, as might be the case if technologists, physician extenders, or radiologists in training, in a core lab or academic centers may be providing these measurements.
Conclusion
This technique of CE-MR imaging with semi-automated volumetric segmentation of necrotic tumor allows non-invasive depiction of the entire tumor and can be performed early in the course of therapy and serially throughout therapy. This may allow for earlier identification of inferior responders, giving clinicians the opportunity to identify a population of patients that may benefit from a change in therapy type or dose. This initial validation suggests the technique to be reproducible, and accurate, correlating with histopathological evaluation.
- Received September 14, 2012.
- Accepted September 25, 2012.
- Copyright© 2012 International Institute of Anticancer Research (Dr. John G. Delinassios), All rights reserved