Lejla Alic

Chapter 1 provides a general introduction to this thesis. The epidemiology of cancer, the biological background of intra-tumour heterogeneity, radiological imaging methods and the methods available for the quantification of heterogeneity are briefly discussed, and the main aim of the thesis is outlined.

Chapter 2 presents a methodology for obtaining an accurate 3D relation between high resolution in vivo T2*-w MRI and the corresponding 3D histology of the tumour tissue. The key features of the methodology are: 1) standardized acquisition and processing, 2) use of an intermediate ex vivo MRI, 3) use of a reference cutting plane, 4) dense histological sampling, 5) elastic registration, and 6) use of complete 3D data sets. The methodology consists of two separate registration steps, both exploiting a three-step strategy of gradually increasing degrees of freedom (rigid, affine, and elastic transformation). These two registration steps involve in vivo MRI to ex vivo MRI registration, and ex vivo MRI to histology registration. The resulting accuracy was assessed by two independent observers and was on average 0.7 mm, between in vivo MRI and 3D reconstructed histology. This accuracy corresponds (on average) with 30-50 cells and is similar to the inter-observer variation of the manual annotations. We have shown that, based on T2*-w MRI signal intensity, automatic identification of necrotic tissue is feasible. However, the separation of two other tissue types, i.e. haemorrhagic and viable tissue, was not possible. For the separation of these tissue types, other MRI sequences are needed. This work is a first step in MRI tumour characterization. Now that spatial correspondence between in vivo MRI and 3D HCE histology has been established, the extension to multi-spectral MR and multi-stained histological sections is the next logical step. The 3D correspondence of tumor histology and in vivo MRI enables extraction of MRI signatures for histologically defined regions.

Chapter 3 provides a systematic review of the literature on tumour heterogeneity quantification methods computed from radiological images for grading, differentiation, outcome prediction and tumour-response monitoring. The heterogeneity analysis methods are divided into four categories, i.e., non-spatial methods (NSM), spatial grey level methods (SGLM), fractal analysis (FA) methods, and filter and transforms (FCT). The reported performance of the heterogeneity features in terms of measures such as accuracy, sensitivity, specificity and AUC, or by statistical tests reporting significance levels, was compared. We found that, among the 8,956 unique studies identified, 192 studies reported heterogeneity as a biomarker in cancer imaging. Since 2009, the number of publications reporting the quantification of tumour heterogeneity has increased. Up until 2006 most heterogeneity papers were based on ultrasound (US), whereas after 2007 the number of studies using MRI has increased. The NSM and the SGLM are the two methods most frequently used during that period. Most papers focus on heterogeneity quantification for differentiation, grading, or outcome prediction has recently increased. The performance of the heterogeneity features was mostly (68%) evaluated by classification experiments reporting performance measures such as accuracy, sensitivity, specificity and AUC. The other papers used only statistical tests (30%) or did not perform a quantitative study (2%). Almost 98% of the studies reported positive findings. The reported AUC ranged from 0.5 to 1 with a median of 0.89. No relation was found between the performance measures and the imaging method or quantification method used. Of all the classification publications, 41% did not report the use of cross-validation as a technique to limit the effect of overfitting on the available data. A negative correlation was found between the tumour-feature ratio and the AUC, which might be caused by overfitting. Publications reporting significance testing often did not perform a correction of the significance levels for multiple comparisons. For 36% of the papers a significant decrease in the number of significant features was observed after the Holm-Bonferroni correction. Only 12% of all included studies had a prospective study design. Although the use of retrospectively collected data is necessary to develop, test and evaluate heterogeneity as a biomarker for tumour grading, differentiation, outcome prediction and treatment response monitoring the real test is to evaluate the performance of the developed features in a prospective study design. Moreover, in most retrospective studies the performance of the heterogeneity feature is evaluated without taking into account other available clinical predictive information. While the researcher may be interested in the performance of the feature itself, the clinician is interested in the additional value of the feature on top of the already available clinical features. To enable the translation of imaging biomarkers from the research stage to clinical practice, future research should focus on prospective studies to investigate the additional value of the proposed heterogeneity biomarker on top of the clinically established markers.

Chapter 4 uses the DCE-MRI images of sarcoma patients undergoing isolated limb perfusion (ILP) with TNF- and melphalan, to evaluate promising imaging-biomarkers: enhancing fraction, SGLM and FA. The potential of these imaging biomarkers to monitor tumour changes and predict tumour response to treatment was evaluated in 18 patients. Using routinely acquired DCE-MRI scans, this study investigates the potential of SGLM and FA to measure the treatment-induced tumour changes for different pharmacokinetic modelling approaches. The correlation between pharmacokinetic and different heuristic parametric maps, averaged over all tumours, was estimated by Spearman’s rank correlation coefficient and ranged between 0.44 and 0.61. The monitoring study demonstrated that, regardless of the origin of the estimated parametric map (pharmacokinetic or heuristic model-based), the enhancing fraction and coherence showed a significant difference between baseline and follow-up acquisitions for the group that responded to therapy. When evaluating the enhancing fraction and coherence on the pre-treatment scan, it was found that both features differentiate between outcome categories. It seems that patients with a large viable tumour fraction and a high coherence respond well to therapy, whereas patients with large necrotic areas and a low coherence show less response to treatment. Moreover, we demonstrated that heterogeneity measures derived from DCE-MRI parametric maps (irrespective of the exact nature of the parametric map) were related to tumour response to chemotherapy. Therefore, we can conclude that DCE-MRI has potential for in vivo monitoring of the tumour during chemotherapy. When correlated to histopathological findings, this method may be clinically useful in understanding pathophysiological changes, predicting tumour response and guiding the therapeutic approach.

Chapter 5 investigates the regional heterogeneity changes in DCE-MRI as response to ILP in an experimental soft-tissue sarcoma model. DCE-MRI of drug-treated and sham-treated rats was performed at baseline and 1 h after ILP intervention. The enhancing data were acquired using a macromolecular contrast medium (MMCM), albumin-(Gd-DTPA) 45. To pinpoint the regional changes accurately, the DCE-MRI at baseline and follow-up were spatially registered. To assess the regional heterogeneity, tumours were divided into 16 sectors. For each sector the cumulative map-volume (CMV) histogram of K trans was computed and the variance in its slope was used as a measure for tumour heterogeneity. The results indicate that the heterogeneity between sectors decreases due to treatment. This implies that the treatment induces tumour homogenization with respect to K trans demonstrating the potential of regional analysis for evaluating local treatment effects of ILP intervention. When part of the tumour ‘escapes’ from treatment, this can be detrimental for the whole treatment effect. Using a regional analysis this ‘escape’ could potentially be spotted within hours after the start of treatment. Therefore, CMVs may serve as non-invasive biomarker of early treatment effects and may guide therapy adaptation.