Presurgical Identification of Primary Central Nervous System Lymphoma with Normalized Time-Intensity Curve: A Pilot Study of a New Method to Analyze DSC-PWI

DISCUSSION

In this study, we have reported the design of an innovative method to obtain normalized TICs from DSC-PWI beyond patient and technical differences, which allows the following: 1) constructing mean curves for visual analysis, 2) performing point-by-point statistical comparisons between curves, and 3) building classifiers. We have tested its applicability in the presurgical identification of PCNSL and obtained satisfactory results.

DSC-PWI is an MR imaging technique that can be performed on most MR imaging units currently and provides noninvasive in vivo assessment of microvascular systems. It consists of a dynamic temporal acquisition during the vascular first pass of a contrast bolus. The injection of gadolinium results in an initial reduction in T2 signal intensity of tissues and subsequent signal recovery during contrast washout. TICs can be generated from this process. Well-studied parameters such as rCBV and PSR are extracted from these curves. The rCBV corresponds to the AUC, is usually measured relative to the NAWM, and has been related to histologic measurements of tumor vascularization.^22,24,30 The PSR is measured relative to the TIC baseline and may quantify the predominant T1 (signal recovery above baseline) or T2 (signal recovery below baseline) effects. These effects represent different leakage phenomena, which are explained by a complex combination of blood-brain barrier permeability, vascular volume fraction and vessel size, and tumor cell size and density.^22,30,31 The extraction of rCBV or PSR from TICs may represent an oversimplification of the information contained in the entire TIC. In fact, the curves have many other points that remain excluded from these parameters. Along this line, some studies have suggested analyzing the whole curve to obtain improved information. Unfortunately, the proposed analysis was qualitative and limited to the visual pattern evaluation of the curve.²¹ Quantitative assessment of the entire curve has not been accomplished to date, to our knowledge. This could be due to differences in the acquisition technique (including operator-dependency on some parameters) and patient physiologic features, which produce noncomparable TICs between different examinations or patients.^22,23 For example, there may be differences in the number and time of dynamics, in contrast injection start point and speed, or in patient heart rate preventing TICs from being comparable.^16,22⇓-24

With these considerations in mind, we have developed a novel method for obtaining standardized, normalized, and comparable TICs independent of some technical and patient variability. We considered 2 parameters to be normalized to obtain comparable curves: time and intensity signal. Time was normalized as constant proportions of TTP-TTA from NAWM, and the time axis was re-dimensioned from seconds to fractions of 0.2 TTP-TTA units. Signal intensity was normalized to the MSID in NAWM. This normalization approach provided superimposable curves that could be visually analyzed and a list of point values that could be statistically compared among cases. Indeed, after normalization, mean curves for each tumor group could be constructed, allowing visual comparisons; the best discriminatory points with their optimal weighting for discrimination could be statistically determined, enabling the construction of classifiers; and particular cases could be displayed on scatterplots, providing visual representations of the likelihood of the diagnostic classification. Accordingly, we consider that this methodology could be further applied to construct user-friendly classifiers for the diagnosis of brain tumors. Examples of this potential application are shown in Figs 3 and 4.

We tested the performance of our method in the presurgical identification of PCNSL.

Reliable presurgical identification of PCNSL is vital because its management greatly differs from that of the other most prevalent enhancing brain tumors.^1⇓-3,32 Maximal PCNSL resection is not recommended, and early stereotactic biopsy before corticosteroid administration is mandatory when it is suspected from imaging.^32⇓⇓-35 Conventional MR imaging in PCNSL has been widely analyzed and may be useful for guiding initial management.^5⇓-7 Nevertheless, these features may vary between patients and may overlap with other tumors.^8⇓⇓-11 Thus, radiologic diagnosis of PCNSL remains a challenge, and additional advanced imaging techniques such as DSC-PWI are increasingly being evaluated. Many articles have evaluated the potential of DSC-PWI for differentiating PCNSL from other tumors with excellent results. These studies focus on rCBV and PSR quantifications. Basically, PCNSL shows low rCBV and high PSR.^{12⇓⇓⇓⇓⇓⇓⇓-20} Some authors recently reported an additional parameter directly extracted from TICs termed “peak height,” which has shown promising results.^12,15 However, the lack of technique standardization, which causes variability in the identification of the best discriminating parameter and its relevant thresholds between different studies, as well as the lack of a user-friendly way to depict the results, impedes the widespread clinical application of these perfusion parameters.^16,22⇓-24

We obtained satisfactory accuracy values in the comparisons between PCNSL and glioblastoma, anaplastic astrocytoma, metastasis, and meningioma in the test subset. Accuracies ranged between 71% (versus anaplastic astrocytoma) and 96% (versus meningioma). Moreover, the performance of the new method is overall superior compared with the analysis of conventional rCBV or PSR measures in our dataset (On-line Tables 6 and 7). Visual differences between the standardized nTICs of the different tumors were noted, especially in the segments around the MSID and the return to baseline (Fig 2). Statistical analysis confirmed that the best discriminatory points were situated around those segments of the curves that may somehow be related to the traditional rCBV, peak height, and PSR, which can be evaluated on conventional raw TICs. Indeed, we hypothesize that our method evaluates a mixture of these known relevant values along with other potentially discriminatory and otherwise hidden values of the curve, all together in a single step. Additionally, the method enables a user-friendly representation of the results (Figs 3 and 4). For this, we used a pair-wise model that takes advantage of the radiologist’s interaction by narrowing the most probable diagnoses. Then, the classifier is used as a support tool for diagnosis and not as an independent reader.

The variety of DSC pulse sequence parameters included in this retrospective study deserves special attention. Differences on these parameters (flip angle, TE, TR) affect the curve morphology and indeed seem to partially justify the variability in values and thresholds found in the literature regarding both rCBV and PSR.³¹ In this sense, an overall predominantly high T1-weighting of the sequences in our study (On-line Table 1) seems to be carrying higher PSR values if compared with some prior studies.²⁰

Several limitations of our study must be considered. The single-site and retrospective character of the study may affect reproducibility. Nevertheless, the single-site origin may confer homogeneity that could be useful for this pilot study. At any rate, multicentric and prospective studies in new real clinical scenarios are needed for validation. The inclusion of a wide range of MR imaging examination dates and consequent technical differences may have affected the consistency of results. Some of them, such as the timing of dynamics or little heterogeneities in technical parameters, may be considered positive for the study by demonstrating the robustness of the method. Other parameters, such as variations in TE, TR, or flip angle, may be considered potentially confusing. In particular, higher T1-weighting of older sequences in our study may overestimate PSR values.^20,31 Balancing of cases between tumor types and training and test cohorts was an attempt to minimize its impact on the results. Finally, we did not stratify subtypes of tumors or DSC sequence parameters used to avoid excessive fragmentation of the dataset.

On the other hand, our study has several strong points. First, a large sample of PCNSL was included, which provides a robust method and very high accuracy rates despite heterogeneity. Second, a wide-but-logical differential diagnosis was considered in the comparisons, which emphasizes the clinical usefulness of the results. Third, there is the semiautomatization of the image segmentation and coregistration as well as the TIC processing, which minimized operator-dependency in favor of reproducibility. Last, the method allowed including all the nTIC point values in the analysis without supervision, which provides an objective approach to the classification process.