Frank Thiele

The quantification of dynamic FET PET imaging and correlation with the clinical outcome in patients with glioblastoma

The PET tracer O-(2-[18F]Fluoroethyl)-l-tyrosine (FET) has been shown to be valuable for different roles in the management of brain tumours. The aim of this study was to evaluate several quantitative measures of dynamic FET PET imaging in patients with resected glioblastoma. We evaluated dynamic FET PET in nine patients with histologically confirmed glioblastoma. Following FET PET, all subjects had radiation and chemotherapy. Tumour ROIs were defined by a threshold-based region-growing algorithm. We compared several standard measures of tumour uptake and uptake kinetics: SUV, SUV/background, distribution volume ratio (DVR), weighted frame differences and compartment model parameters. These measures were correlated with disease-free and overall survival, and analysed for statistical significance. We found that several measures allowed robust quantification. SUV and distribution volume did not correlate with clinical outcome. Measures that are based on a background region (SUV/BG, Logan-DVR) highly correlated with disease-free survival (r = -0.95, p < 0.0001), but not overall survival. Some advanced measures also showed a prognostic value but no improvement over the simpler methods. We conclude that FET PET probably has a prognostic value in patients with resected glioblastoma. The ratio of SUV to background may provide a simple and valuable predictive measure of the clinical outcome. Further studies are needed to confirm these explorative results.

Evaluation of non-uniform weighting in non-linear regression for pharmacokinetic neuroreceptor modelling.

ObjectivePharmacokinetic modelling of dynamic PET data has become an important tool to analyse in-vivo studies in humans and animals. Estimation of the model parameters often requires non-linear regression of an objective function such as weighted least squares. Since the noise properties of the data are not known exactly in practice, several weighting schemes have been proposed. The objective of this study was to evaluate the impact of commonly used weights on neuroreceptor quantification with the simplified reference tissue model (SRTM). MethodsWe compared the following weights: uniform, Poisson statistics-based ideal and noisy weights, iterative weighting, and a noise-free approximation of Poisson weights. Ten thousand time–activity curves (TACs) were simulated for several noise levels and the three neuroreceptor PET ligands 11C-(+)McN5652, 11C-DASB, and 11C-raclopride. Each TAC was fitted using weighted non-linear regression of the SRTM. We assessed bias and variation of the parameter estimates as well as quality of fit and parameter distributions. ResultsResults differed substantially between ligands and between model parameters. Best parameter estimates were obtained with the noise-free approximation of Poisson weights. The often-used noisy Poisson weights performed worst for all ligands. Uniform weighting gave acceptable parameter estimates for most setups. Conclusion‘Choice of weights’ is important in pharmacokinetic neuroreceptor quantification with the SRTM. Weights estimated directly from noisy data should be avoided as they can severely degrade parameter estimation and the statistical power of a study. If the noise characteristic of the data is unknown, uniform weighting is recommended.

Long range language models for free spelling recognition

Highly accurate spelling recognizers are essential in many commercially relevant applications. Examples are directory assistance, address taking in ordering services or help desks, and input of difficult or unknown words in dictation. Language models whose context length is flexibly configured by incorporating automatically determined letter groups code the structure of the recognition items in traditional bi- or trigrams. This gives a powerful method trading off computational demands versus recognition accuracy, comparing favorably with the standard word-list constraint. In addition, the automatic modeling of new words proves its benefits in applications with high out-of-vocabulary rates. For the task of spelling German last names over the telephone, letter error rates could be improved from 12.5% using a standard bigram to 3.6% with a trigram on a set of 2782 letter groups, giving the additional benefit of recognizing about 40% of the names not seen before in the language model training corpus.

Construction and comparative evaluation of different activity detection methods in brain FDG‑PET

AimWe constructed and evaluated reference brain FDG-PET databases for usage by three software programs (Computer-aided diagnosis for dementia (CAD4D), Statistical Parametric Mapping (SPM) and NEUROSTAT), which allow a user-independent detection of dementia-related hypometabolism in patients’ brain FDG-PET.MethodsThirty-seven healthy volunteers were scanned in order to construct brain FDG reference databases, which reflect the normal, age-dependent glucose consumption in human brain, using either software. Databases were compared to each other to assess the impact of different stereotactic normalization algorithms used by either software package. In addition, performance of the new reference databases in the detection of altered glucose consumption in the brains of patients was evaluated by calculating statistical maps of regional hypometabolism in FDG-PET of 20 patients with confirmed Alzheimer’s dementia (AD) and of 10 non-AD patients. Extent (hypometabolic volume referred to as cluster size) and magnitude (peak z-score) of detected hypometabolism was statistically analyzed.ResultsDifferences between the reference databases built by CAD4D, SPM or NEUROSTAT were observed. Due to the different normalization methods, altered spatial FDG patterns were found. When analyzing patient data with the reference databases created using CAD4D, SPM or NEUROSTAT, similar characteristic clusters of hypometabolism in the same brain regions were found in the AD group with either software. However, larger z-scores were observed with CAD4D and NEUROSTAT than those reported by SPM. Better concordance with CAD4D and NEUROSTAT was achieved using the spatially normalized images of SPM and an independent z-score calculation. The three software packages identified the peak z-scores in the same brain region in 11 of 20 AD cases, and there was concordance between CAD4D and SPM in 16 AD subjects.ConclusionThe clinical evaluation of brain FDG-PET of 20 AD patients with either CAD4D-, SPM- or NEUROSTAT-generated databases from an identical reference dataset showed similar patterns of hypometabolism in the brain regions known to be involved in AD. The extent of hypometabolism and peak z-score appeared to be influenced by the calculation method used in each software package rather than by different spatial normalization parameters.

Optimal feature selection for automated classification of FDG-PET in patients with suspected dementia

FDG-PET is increasingly used for the evaluation of dementia patients, as major neurodegenerative disorders, such as Alzheimers disease (AD), Lewy body dementia (LBD), and Frontotemporal dementia (FTD), have been shown to induce specific patterns of regional hypo-metabolism. However, the interpretation of FDG-PET images of patients with suspected dementia is not straightforward, since patients are imaged at different stages of progression of neurodegenerative disease, and the indications of reduced metabolism due to neurodegenerative disease appear slowly over time. Furthermore, different diseases can cause rather similar patterns of hypo-metabolism. Therefore, classification of FDG-PET images of patients with suspected dementia may lead to misdiagnosis. This work aims to find an optimal subset of features for automated classification, in order to improve classification accuracy of FDG-PET images in patients with suspected dementia. A novel feature selection method is proposed, and performance is compared to existing methods. The proposed approach adopts a combination of balanced class distributions and feature selection methods. This is demonstrated to provide high classification accuracy for classification of FDG-PET brain images of normal controls and dementia patients, comparable with alternative approaches, and provides a compact set of features selected.

Combining Image and Non-image Clinical Data: An Infrastructure that Allows Machine Learning Studies in a Hospital Environment

Over the past years Machine Learning and Deep Learning techniques are showing their huge potential in medical research. However, this research is mainly done by using public or private datasets that were created for study purposes. Despite ensuring reproducibility, these datasets need to be constantly updated.

Fully automated detection and segmentation of meningiomas using deep learning on routine multiparametric MRI

ObjectivesMagnetic resonance imaging (MRI) is the method of choice for imaging meningiomas. Volumetric assessment of meningiomas is highly relevant for therapy planning and monitoring. We used a multiparametric deep-learning model (DLM) on routine MRI data including images from diverse referring institutions to investigate DLM performance in automated detection and segmentation of meningiomas in comparison to manual segmentations.MethodsWe included 56 of 136 consecutive preoperative MRI datasets [T1/T2-weighted, T1-weighted contrast-enhanced (T1CE), FLAIR] of meningiomas that were treated surgically at the University Hospital Cologne and graded histologically as tumour grade I (n = 38) or grade II (n = 18). The DLM was trained on an independent dataset of 249 glioma cases and segmented different tumour classes as defined in the brain tumour image segmentation benchmark (BRATS benchmark). The DLM was based on the DeepMedic architecture. Results were compared to manual segmentations by two radiologists in a consensus reading in FLAIR and T1CE.ResultsThe DLM detected meningiomas in 55 of 56 cases. Further, automated segmentations correlated strongly with manual segmentations: average Dice coefficients were 0.81 ± 0.10 (range, 0.46-0.93) for the total tumour volume (union of tumour volume in FLAIR and T1CE) and 0.78 ± 0.19 (range, 0.27-0.95) for contrast-enhancing tumour volume in T1CE.ConclusionsThe DLM yielded accurate automated detection and segmentation of meningioma tissue despite diverse scanner data and thereby may improve and facilitate therapy planning as well as monitoring of this highly frequent tumour entity.Key Points• Deep learning allows for accurate meningioma detection and segmentation• Deep learning helps clinicians to assess patients with meningiomas• Meningioma monitoring and treatment planning can be improved