Spyridon Bakas

Advancing The Cancer Genome Atlas glioma MRI collections with expert segmentation labels and radiomic features

Gliomas belong to a group of central nervous system tumors, and consist of various sub-regions. Gold standard labeling of these sub-regions in radiographic imaging is essential for both clinical and computational studies, including radiomic and radiogenomic analyses. Towards this end, we release segmentation labels and radiomic features for all pre-operative multimodal magnetic resonance imaging (MRI) (n=243) of the multi-institutional glioma collections of The Cancer Genome Atlas (TCGA), publicly available in The Cancer Imaging Archive (TCIA). Pre-operative scans were identified in both glioblastoma (TCGA-GBM, n=135) and low-grade-glioma (TCGA-LGG, n=108) collections via radiological assessment. The glioma sub-region labels were produced by an automated state-of-the-art method and manually revised by an expert board-certified neuroradiologist. An extensive panel of radiomic features was extracted based on the manually-revised labels. This set of labels and features should enable i) direct utilization of the TCGA/TCIA glioma collections towards repeatable, reproducible and comparative quantitative studies leading to new predictive, prognostic, and diagnostic assessments, as well as ii) performance evaluation of computer-aided segmentation methods, and comparison to our state-of-the-art method.

Fast Semi-Automatic Segmentation of Focal Liver Lesions in Contrast-Enhanced Ultrasound, based on a Probabilistic Model

Assessment of focal liver lesions (FLLs) in contrast-enhanced ultrasound requires the delineation of the FLL in at least one frame of the acquired data, which is currently performed manually by experienced radiologists. Such a task leads to subjective results, is time-consuming and prone to misinterpretation and human error. This paper describes an attempt to improve this clinical practice by proposing a novel fast two-step method to automate the FLL segmentation, initialised only by a single seed point. Firstly, rectangular force functions are used to improve the accuracy and computational efficiency of an active ellipse model for approximating the FLL shape. Then, a novel probabilistic boundary refinement method is used to iteratively classify boundary pixels rapidly. The proposed method allows for faster and easier assessment of FLLs, whilst requiring less interaction, but producing results comparably consistent with manual delineations, and hence increasing the confidence of radiologists when making a diagnosis. Quantitative evaluation based on real clinical data, from two different European countries reflecting true clinical practice, demonstrates the value of the proposed method.

GLISTRboost: Combining Multimodal MRI Segmentation, Registration, and Biophysical Tumor Growth Modeling with Gradient Boosting Machines for Glioma Segmentation

We present an approach for segmenting low- and high-grade gliomas in multimodal magnetic resonance imaging volumes. The proposed approach is based on a hybrid generative-discriminative model. Firstly, a generative approach based on an Expectation-Maximization framework that incorporates a glioma growth model is used to segment the brain scans into tumor, as well as healthy tissue labels. Secondly, a gradient boosting multi-class classification scheme is used to refine tumor labels based on information from multiple patients. Lastly, a probabilistic Bayesian strategy is employed to further refine and finalize the tumor segmentation based on patient-specific intensity statistics from the multiple modalities. We evaluated our approach in 186 cases during the training phase of the BRAin Tumor Segmentation (BRATS) 2015 challenge and report promising results. During the testing phase, the algorithm was additionally evaluated in 53 unseen cases, achieving the best performance among the competing methods.

Spot the Best Frame: Towards Intelligent Automated Selection of the Optimal Frame for Initialisation of Focal Liver Lesion Candidates in Contrast-Enhanced Ultrasound Video Sequences

This paper describes a contribution to a wider project which aims to provide an intelligent automated assistant to radiologists performing the skilled and time-intensive task of detecting and characterising cancerous lesions within a human liver from Contrast-Enhanced Ultrasound (CEUS) video sequences. This particular contribution relates to automatically locating the optimal frame, for initialising a suspected focal liver lesion (FLL), within a CEUS video sequence. Currently, this task is routinely performed manually by radiologists, but is very time-consuming. The proposed approach is to use statistical and image processing techniques to automatically identify the most suitable frame for performing this initialisation, which should save the radiologist significant time and effort, bearingin mind the continuously increasing amount of CEUS data acquired and processed. In the future, this could be coupled with a method for automatically initialising the FLLs area within the area of the ultrasonographic image in this optimal frame and, together with already produced systems for the tracking and characterisation of such lesions, lead to a fully automated system assisting clinicians in the diagnosis of such lesions.

Focal Liver Lesion Tracking in CEUS for Characterisation Based on Dynamic Behaviour

This paper presents a methodology for tracking a hypo- or hyper-enhanced focal liver lesion (FLL) and a healthy liver region in a video sequence of a Contrast-Enhanced Ultrasound (CEUS) examination. The outcome allows the differentiation between benign and malignant cases, by characterising FLLs of typical behaviour, according to their Time-Intensity curves. The task is challenging mainly due to intensity changes caused by contrast agents. Initially the ultrasound mask is automatically localised and then the FLL and parenchyma regions are tracked, assuming affine transformations on the image plane, employing the point-based registration technique of Lowe’s scale-invariant feature transform (SIFT) keypoints detector. Finally, a quantitative evaluation of the tracking process provides a confidence measure for the characterisation decision.

In vivo detection of EGFRvIII in glioblastoma via perfusion magnetic resonance imaging signature consistent with deep peritumoral infiltration: the φ index

Purpose: The epidermal growth factor receptor variant III (EGFRvIII) mutation has been considered a driver mutation and therapeutic target in glioblastoma, the most common and aggressive brain cancer. Currently, detecting EGFRvIII requires postoperative tissue analyses, which are ex vivo and unable to capture the tumors spatial heterogeneity. Considering the increasing evidence of in vivo imaging signatures capturing molecular characteristics of cancer, this study aims to detect EGFRvIII in primary glioblastoma noninvasively, using routine clinically acquired imaging. Experimental Design: We found peritumoral infiltration and vascularization patterns being related to EGFRvIII status. We therefore constructed a quantitative within-patient peritumoral heterogeneity index (PHI/ϕ-index), by contrasting perfusion patterns of immediate and distant peritumoral edema. Application of ϕ-index in preoperative perfusion scans of independent discovery (n = 64) and validation (n = 78) cohorts, revealed the generalizability of this EGFRvIII imaging signature. Results: Analysis in both cohorts demonstrated that the obtained signature is highly accurate (89.92%), specific (92.35%), and sensitive (83.77%), with significantly distinctive ability (P = 4.0033 × 10−10, AUC = 0.8869). Findings indicated a highly infiltrative-migratory phenotype for EGFRvIII+ tumors, which displayed similar perfusion patterns throughout peritumoral edema. Contrarily, EGFRvIII− tumors displayed perfusion dynamics consistent with peritumorally confined vascularization, suggesting potential benefit from extensive peritumoral resection/radiation. Conclusions: This EGFRvIII signature is potentially suitable for clinical translation, since obtained from analysis of clinically acquired images. Use of within-patient heterogeneity measures, rather than population-based associations, renders ϕ-index potentially resistant to inter-scanner variations. Overall, our findings enable noninvasive evaluation of EGFRvIII for patient selection for targeted therapy, stratification into clinical trials, personalized treatment planning, and potentially treatment-response evaluation. Clin Cancer Res; 23(16); 4724–34. ©2017 AACR.

Automatic Identification and Localisation of Potential Malignancies in Contrast-Enhanced Ultrasound Liver Scans Using Spatio-Temporal Features

The identification and localisation of a focal liver lesion (FLL) in Contrast-Enhanced Ultrasound (CEUS) video sequences is crucial for liver cancer diagnosis, treatment planning and follow-up management. Currently, localisation and classification of FLLs between benign and malignant cases in CEUS are routinely performed manually by radiologists, in order to proceed with making a diagnosis, leading to subjective results, prone to misinterpretation and human error. This paper describes a methodology to assist clinicians who regularly perform these tasks, by discharging benign FLL cases and localise potential malignancies in a fully automatic manner by exploiting the perfusion dynamics of a CEUS video. The proposed framework uses local variations of intensity to distinguish between hyper- and hypo-enhancing regions and then analyse their spatial configuration to identify potentially malignant cases. Automatic localisation of the potential malignancy on the image plane is then addressed by clustering, using Expectation-Maximisation for Gaussian Mixture Models. A novel feature that combines description of local dynamic behaviour with spatial proximity is used in this process. Quantitative evaluation, on real clinical data from a retrospective multi-centre study, demonstrates the value of the proposed method.

Cancer imaging phenomics toolkit: quantitative imaging analytics for precision diagnostics and predictive modeling of clinical outcome

Abstract. The growth of multiparametric imaging protocols has paved the way for quantitative imaging phenotypes that predict treatment response and clinical outcome, reflect underlying cancer molecular characteristics and spatiotemporal heterogeneity, and can guide personalized treatment planning. This growth has underlined the need for efficient quantitative analytics to derive high-dimensional imaging signatures of diagnostic and predictive value in this emerging era of integrated precision diagnostics. This paper presents cancer imaging phenomics toolkit (CaPTk), a new and dynamically growing software platform for analysis of radiographic images of cancer, currently focusing on brain, breast, and lung cancer. CaPTk leverages the value of quantitative imaging analytics along with machine learning to derive phenotypic imaging signatures, based on two-level functionality. First, image analysis algorithms are used to extract comprehensive panels of diverse and complementary features, such as multiparametric intensity histogram distributions, texture, shape, kinetics, connectomics, and spatial patterns. At the second level, these quantitative imaging signatures are fed into multivariate machine learning models to produce diagnostic, prognostic, and predictive biomarkers. Results from clinical studies in three areas are shown: (i) computational neuro-oncology of brain gliomas for precision diagnostics, prediction of outcome, and treatment planning; (ii) prediction of treatment response for breast and lung cancer, and (iii) risk assessment for breast cancer.

Segmentation of gliomas in pre-operative and post-operative multimodal magnetic resonance imaging volumes based on a hybrid generative-discriminative framework

We present an approach for segmenting both low- and high-grade gliomas in multimodal magnetic resonance imaging volumes. The proposed framework is an extension of our previous work [6,7], with an additional component for segmenting post-operative scans. The proposed approach is based on a hybrid generative-discriminative model. Firstly, a generative model based on a joint segmentation-registration framework is used to segment the brain scans into cancerous and healthy tissues. Secondly, a gradient boosting classification scheme is used to refine tumor segmentation based on information from multiple patients. We evaluated our approach in 218 cases during the training phase of the BRAin Tumor Segmentation (BRATS) 2016 challenge and report promising results. During the testing phase, the proposed approach was ranked among the top performing methods, after being additionally evaluated in 191 unseen cases.

Correlations of atrial diameter and frontooccipital horn ratio with ventricle size in fetal ventriculomegaly

OBJECTIVE Fetal ventriculomegaly (FV), or enlarged cerebral ventricles in utero, is defined in fetal studies as an atrial diameter (AD) greater than 10 mm. In postnatal studies, the frontooccipital horn ratio (FOHR) is commonly used as a proxy for ventricle size (VS); however, its role in FV has not been assessed. Using image analysis techniques to quantify VS on fetal MR images, authors of the present study examined correlations between linear measures (AD and FOHR) and VS in patients with FV. METHODS The authors performed a cross-sectional study using fetal MR images to measure AD in the axial plane at the level of the atria of the lateral ventricles and to calculate FOHR as the average of the frontal and occipital horn diameters divided by the biparietal distance. Computer software was used to separately segment and measure the area of the ventricle and the ventricle plus the subarachnoid space in 2 dimensions. Segmentation was performed on axial slices 3 above and 3 below the slice used to measure AD, and measurements for each slice were combined to yield a volume, or 3D VS. The VS was expressed as the absolute number of voxels (non-normalized) and as the number of voxels divided by intracranial size (normalized). A Pearson correlation coefficient was used to measure the strength of the relationships between the linear measures and the size of segmented regions in 2 and 3 dimensions and over various gestational ages (GAs). Differences between correlations were compared using Steigers z-test. RESULTS Fifty FV patients who had undergone fetal MRI between 2008 and 2014 were included in the study. The mean GA was 26.3 ± 5.4 weeks. The mean AD was 18.1 ± 8.3 mm, and the mean FOHR was 0.49 ± 0.11. When using absolute VS, the correlation between AD and 3D VS (r = 0.844, p < 0.0001) was significantly higher than that between FOHR and 3D VS (r = 0.668, p < 0.0001; p = 0.0004, Steigers z-test). However, when VS was normalized, correlations were not significantly different between AD and 3D VS (r = 0.830, p < 0.0001) or FOHR and 3D VS (r = 0.842, p < 0.0001; p = 0.8, Steigers z-test). For GAs of 24 weeks or earlier, AD correlated more strongly with normalized 3D VS (r = 0.902, p < 0.0001) than with FOHR (r = 0.674, p < 0.0001; p < 0.0001, Steigers z-test). After 24 weeks, there was no difference in correlations between linear measures (AD or FOHR) and 3D VS (r > 0.9). Correlations of linear measures with VS in 2 and 3 dimensions were similar, and inclusion of the subarachnoid space did not significantly alter results. CONCLUSIONS Findings in the study support the use of AD as a measure of VS in fetal studies as it correlates highly with both absolute and relative VS, especially at early GAs, and captures the preferential dilation of the occipital horns in patients with FV. Compared with AD, FOHR similarly correlates with normalized VS and, after a GA of 24 weeks, can be reported in fetal studies to provide continuity with postnatal monitoring.