Automated Meningioma Segmentation in Multiparametric MRI : Comparable Effectiveness of a Deep Learning Model and Manual Segmentation.

PURPOSE: Volumetric assessment of meningiomas represents a valuable tool for treatment planning and evaluation of tumor growth as it enables a more precise assessment of tumor size than conventional diameter methods. This study established a dedicated meningioma deep learning model based on routine magnetic resonance imaging (MRI) data and evaluated its performance for automated tumor segmentation. METHODS: The MRI datasets included T1-weighted/T2-weighted, T1-weighted contrast-enhanced (T1CE) and FLAIR of 126 patients with intracranial meningiomas (grade I: 97, grade II: 29). For automated segmentation, an established deep learning model architecture (3D deep convolutional neural network, DeepMedic, BioMedIA) operating on all four MR sequences was used. Segmentation included the following two components: (i) contrast-enhancing tumor volume in T1CE and (ii) total lesion volume (union of lesion volume in T1CE and FLAIR, including solid tumor parts and surrounding edema). Preprocessing of imaging data included registration, skull stripping, resampling, and normalization. After training of the deep learning model using manual segmentations by 2 independent readers from 70 patients (training group), the algorithm was evaluated on 56 patients (validation group) by comparing automated to ground truth manual segmentations, which were performed by 2 experienced readers in consensus. RESULTS: Of the 56 meningiomas in the validation group 55 were detected by the deep learning model. In these patients the comparison of the deep learning model and manual segmentations revealed average dice coefficients of 0.91 ± 0.08 for contrast-enhancing tumor volume and 0.82 ± 0.12 for total lesion volume. In the training group, interreader variabilities of the 2 manual readers were 0.92 ± 0.07 for contrast-enhancing tumor and 0.88 ± 0.05 for total lesion volume. CONCLUSION: Deep learning-based automated segmentation yielded high segmentation accuracy, comparable to manual interreader variability.

Accuracy of Radiomics-Based Feature Analysis on Multiparametric Magnetic Resonance Images for Noninvasive Meningioma Grading.

OBJECTIVE: Meningioma grading is relevant to therapy decisions in complete or partial resection, observation, and radiotherapy because higher grades are associated with tumor growth and recurrence. The differentiation of low and intermediate grades is particularly challenging. This study attempts to apply radiomics-based shape and texture analysis on routine multiparametric magnetic resonance imaging (MRI) from different scanners and institutions for grading. METHODS: We used MRI data (T1-weighted/T2-weighted, T1-weighted-contrast-enhanced [T1CE], fluid-attenuated inversion recovery [FLAIR], diffusion-weighted imaging [DWI], apparent diffusion coefficient [ADC]) of grade I (n = 46) and grade II (n = 25) nontreated meningiomas with histologic workup. Two experienced radiologists performed manual tumor segmentations on FLAIR, T1CE, and ADC images in consensus. The MRI data were preprocessed through T1CE and T1-subtraction, coregistration, resampling, and normalization. A PyRadiomics package was used to generate 990 shape/texture features. Stepwise dimension reduction and robust radiomics feature selection were performed. Biopsy results were used as standard of reference. RESULTS: Four statistically independent radiomics features were identified as showing the strongest predictive values for higher tumor grades: roundness-of-FLAIR-shape (area under curve [AUC], 0.80), cluster-shades-of-FLAIR/T1CE-gray-level (AUC, 0.80), DWI/ADC-gray-level-variability (AUC, 0.72), and FLAIR/T1CE-gray-level-energy (AUC, 0.76). In a multivariate logistic regression model, the combination of the features led to an AUC of 0.91 for the differentiation of grade I and grade II meningiomas. CONCLUSIONS: Our results indicate that radiomics-based feature analysis applied on routine MRI is viable for meningioma grading, and a multivariate logistic regression model yielded strong classification performances. More advanced tumor stages are identifiable through certain shape parameters of the lesion, textural patterns in morphologic MRI sequences, and DWI/ADC variability.

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

OBJECTIVES: Magnetic 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. METHODS: We 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. RESULTS: The 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. CONCLUSIONS: The 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.

Clinical Evaluation of a Multiparametric Deep Learning Model for Glioblastoma Segmentation Using Heterogeneous Magnetic Resonance Imaging Data From Clinical Routine.

OBJECTIVES: The aims of this study were, first, to evaluate a deep learning-based, automatic glioblastoma (GB) tumor segmentation algorithm on clinical routine data from multiple centers and compare the results to a ground truth, manual expert segmentation, and second, to evaluate the quality of the segmentation results across heterogeneous acquisition protocols of routinely acquired clinical magnetic resonance imaging (MRI) examinations from multiple centers. MATERIALS AND METHODS: The data consisted of preoperative MRI scans (T1, T2, FLAIR, and contrast-enhanced [CE] T1) of 64 patients with an initial diagnosis of primary GB, which were acquired in 15 institutions with varying protocols. All images underwent preprocessing (coregistration, skull stripping, resampling to isotropic resolution, normalization) and were fed into an independently trained deep learning model based on DeepMedic, a multilayer, multiscale convolutional neural network for detection and segmentation of tumor compartments. Automatic segmentation results for the whole tumor, necrosis, and CE tumor were compared with manual segmentations. RESULTS: Whole tumor and CE tumor compartments were correctly detected in 100% of the cases; necrosis was correctly detected in 91% of the cases. A high segmentation accuracy comparable to interrater variability was achieved for the whole tumor (mean dice similarity coefficient [DSC], 0.86 ± 0.09) and CE tumor (DSC, 0.78 ± 0.15). The DSC for tumor necrosis was 0.62 ± 0.30. We have observed robust segmentation quality over heterogeneous image acquisition protocols, for example, there were no correlations between resolution and segmentation accuracy of the single tumor compartments. Furthermore, no relevant correlation was found between quality of automatic segmentation and volume of interest properties (surface-to-volume ratio and volume). CONCLUSIONS: The proposed approach for automatic segmentation of GB proved to be robust on routine clinical data and showed on all tumor compartments a high automatic detection rate and a high accuracy, comparable to interrater variability. Further work on improvements of the segmentation accuracy for the necrosis compartments should be guided by the evaluation of the clinical relevance.Therefore, we propose this approach as a suitable building block for automatic tumor segmentation to support radiologists or neurosurgeons in the preoperative reading of GB MRI images and characterization of primary GB.

A novel concept for tumour targeting with radiation: Inverse dose-painting or targeting the "Low Drug Uptake Volume".

BACKGROUND AND PURPOSE: We tested a novel treatment approach combining (1) targeting radioresistant hypoxic tumour cells with the hypoxia-activated prodrug TH-302 and (2) inverse radiation dose-painting to boost selectively non-hypoxic tumour sub-volumes having no/low drug uptake. MATERIAL AND METHODS: 18F-HX4 hypoxia tracer uptake measured with a clinical PET/CT scanner was used as a surrogate of TH-302 activity in rhabdomyosarcomas growing in immunocompetent rats. Low or high drug uptake volume (LDUV/HDUV) was defined as 40% of the GTV with the lowest or highest 18F-HX4 uptake, respectively. Two hours post TH-302/saline administration, animals received either single dose radiotherapy (RT) uniformly (15 or 18.5Gy) or a dose-painted non-uniform radiation (15Gy) with 50% higher dose to LDUV or HDUV (18.5Gy). Treatment plans were created using Eclipse treatment planning system and radiation was delivered using VMAT. Tumour response was quantified as time to reach 3 times starting tumour volume. RESULTS: Non-uniform RT boosting tumour sub-volume with low TH-302 uptake (LDUV) was superior to the same dose escalation to HDUV (p<0.0001) and uniform RT with the same mean dose 15Gy (p=0.0077). Noteworthy, dose escalation to LDUV required on average 3.5Gy lower dose to the GTV to achieve similar tumour response as uniform dose escalation. CONCLUSIONS: The results support targeted dose escalation to non-hypoxic tumour sub-volume with no/low activity of hypoxia-activated prodrugs. This strategy applies on average a lower radiation dose and is as effective as uniform dose escalation to the entire tumour. It could be applied to other type of drugs provided that their distribution can be imaged.

Liver tissue classification in patients with hepatocellular carcinoma by fusing structured and rotationally invariant context representation.

This work addresses multi-class liver tissue classification from multi-parameter MRI in patients with hepatocellular carcinoma (HCC), and is among the first to do so. We propose a structured prediction framework to simultaneously classify parenchyma, blood vessels, viable tumor tissue, and necrosis, which overcomes limitations related to classifying these tissue classes individually and consecutively. A novel classification framework is introduced, based on the integration of multi-scale shape and appearance features to initiate the classification, which is iteratively refined by augmenting the feature space with both structured and rotationally invariant label context features. We study further the topic of rotationally invariant label context feature representations, and introduce a method for this purpose based on computing the energies of the spherical harmonic decompositions computed at different frequencies and radii. We test our method on full 3D multi-parameter MRI volumes from 47 patients with HCC and achieve promising results.

Probabilistic evaluation of target dose deterioration in dose painting by numbers for stage II/III lung cancer.

PURPOSE: Non-small cell lung cancer is typically irradiated with 60-66 Gy in 2-Gy fractions. Local control could be improved by increasing dose to the more radiation-resistant areas (eg, based on the standardized uptake values of a pretreatment [(18)F]fluoro-deoxyglucose positron emission tomography scan). Such dose painting approaches, however, are poorly suited for a conventional planning target volume margin expansion; therefore, typically no margins are used. This study investigates dose deterioration of a dose painting by numbers (DPBN) approach resulting from geometrical uncertainties. METHODS AND MATERIALS: For 9 DPBN plans of stage II/III non-small cell lung cancer patients, the boost dose was escalated up to 130 Gy (in 33 fractions) or until a dose-limiting constraint was reached. Then, using Monte Carlo methods, a probabilistic evaluation of dose endpoints for 99%, 98%, and 2% of gross tumor volume at a 90% confidence level was performed considering 8 different combinations of systematic (∑) and random (σ) geometric error distributions. RESULTS: Important underdosages, because of geometric uncertainties, of up to 38 Gy with minimal image guidance occur, reducing to 8 Gy with the highest level of image guidance, for a patient where a maximum dose of 119 Gy could be achieved. The evaluation showed that systematic errors had the largest influence. The effects of the uncertainties are most evident where the dose or its gradient is high. CONCLUSIONS: Probabilistic evaluation showed that the geometric uncertainties have a large effect and should be evaluated before approving DPBN plans.

An in silico comparison between margin-based and probabilistic target-planning approaches in head and neck cancer patients.

BACKGROUND AND PURPOSE: To apply target probabilistic planning (TPP) approach to intensity modulated radiotherapy (IMRT) plans for head and neck cancer (HNC) patients. MATERIAL AND METHODS: Twenty plans of HNC patients were re-planned replacing the simultaneous integrated boost IMRT optimization objectives for minimum dose on the boost target and the elective volumes with research probabilistic objectives: the latter allow for explicit handling of systematic and random geometric uncertainties, enabling confidence level based probabilistic treatment planning. Monte-Carlo evaluations of geometrical errors were performed, with endpoints D98%, D2% and Dmean, calculated at a confidence level of 90%. The dose distribution was expanded outside the patient to prevent large bilateral elective treatment volumes ending up in air for probabilistic shifts. RESULTS: TPP resulted in more regular isodoses and in reduced dose, on average, to organs at risk (OAR), up to more than 6Gy, while maintaining target coverage and keeping the maximum dose to limiting structures within requirements. In particular, when the surrounding OARs overlap with the planning target volume (PTV) but not with the clinical target volume (CTV), better results were achieved. CONCLUSION: The TPP approach was evaluated in HNC patients, and proven to be an efficient tool for managing uncertainties.

Dealing with geometric uncertainties in dose painting by numbers: introducing the ΔVH.

PURPOSE: Dose painting by numbers lacks the conventional margin approach for geometric uncertainties. Moreover, the DVH is unable to assess the geometric accuracy of a non-uniform dose distribution because spatial information is lost. In this work we present tools for planning and evaluation of non-uniform treatment dose which take geometric uncertainties into account. METHODS AND MATERIALS: The IMRT optimization functions in the Pinnacle treatment planning software were extended to allow non-uniform prescription dose distributions, e.g., derived from a PET image set. Also, explicit handling of systematic and random geometric uncertainties was incorporated in the functions, enabling confidence level based probabilistic treatment planning. For plan evaluation the concept of ΔVH was introduced, which is the volume histogram of the difference between planned and prescribed doses. Probability distributions for ΔVH points were estimated using Monte Carlo methods. As a demonstration of these methods, two examples are presented; one plan for a lung cancer patient and one for a tumor in the head-and-neck region. RESULTS: Dose distributions were obtained using the PET SUV, while allowing for geometric uncertainties. Optimization was performed such that the ΔVH evaluation indicated a 90% confidence of having under-dosage less than 5% of prescription dose maximum in 99% of the tumor volume. This corresponds to the clinical target constraint for margin based planning with uniform dose prescription. CONCLUSIONS: Clinical treatment planning tools were extended to allow non-uniform prescription. For planning we introduced confidence level based probabilistic optimization with non-uniform target dose, while confidence levels of ΔVH points summarize the probability of proper target coverage.

Implementation and workflow for PET monitoring of therapeutic ion irradiation: a comparison of in-beam, in-room, and off-line techniques.

An independent assessment of the dose delivery in ion therapy can be performed using positron emission tomography (PET). For that a distribution of positron emitters which appear as the result of interaction between ions of the therapeutic beam and the irradiated tissue is measured during or after the irradiation. Three concepts for PET monitoring implemented in various therapy facilities are considered in this paper. The in-beam PET concept relies on the PET measurement performed simultaneously to the irradiation by means of a PET scanner which is completely integrated into the irradiation site. The in-room PET concept allows measurement immediately after irradiation by a standalone PET scanner which is installed very close to the irradiation site. In the off-line PET scenario the measurement is performed by means of a standalone PET/CT scanner 10-30 min after the irradiation. These three concepts were evaluated according to image quality criteria, integration costs, and their influence onto the workflow of radiotherapy. In-beam PET showed the best performance. However, the integration costs were estimated as very high for this modality. Moreover, the performance of in-beam PET depends heavily on type and duty cycle of the accelerator. The in-room PET is proposed for planned therapy facilities as a good compromise between the quality of measured data and integration efforts. For facilities which are close to the nuclear medicine departments off-line PET can be suggested under several circumstances.

Preliminary assessment of dynamic contrast-enhanced CT implementation in pretreatment FDG-PET/CT for outcome prediction in head and neck tumors.

BACKGROUND: Recently published data show some controversy concerning the impact of [18F]-fluorodeoxyglucose positron emission tomography/computed tomography (FDG-PET/CT) in predicting head and neck tumors (HNT) outcome. Assessment of tumor blood supply parameters using dynamic contrast-enhanced CT (DCE-CT) may deliver additional information concerning this important question. PURPOSE: To evaluate the contribution of DCE-CT implemented in pretherapeutic FDG-PET/CT protocol for prognosis prediction in patients with HNT. MATERIAL AND METHODS: Ten consecutive patients (median age 50 years, range 47-74 years) with histologically proven HNT underwent FDG-PET/CT with DCE-CT before treatment. FDG uptake was measured by maximum standardized uptake value (SUV(max)). Relative tumor blood volume (rTBV) was determined from DCE-CT using Patlak analysis. Intratumoral heterogeneity was assessed by means of lacunarity analysis. Obtained values were compared with time-to-progression and overall survival. PET and DCE-CT images were compared on a pixel-by-pixel basis using Pearson coefficient of correlation. RESULTS: Three patients with lower FDG uptake (SUV(max): 8+/-1) and five patients with higher FDG uptake (SUV(max): 15+/-4, P=0.004) were free of local recurrence for 24 months. Two groups of patients with significantly differing lower (group A: 0.37+/-0.02, n=6) and higher (group B: 0.52+/-0.01, n=4; P<0.01), tumor heterogeneity (lacunarity) were identified. Corresponding mean rTBV was higher in group A (9.6+/-1.8 ml/100 ml) than in group B (6.2+/-0.6 ml/100 ml). All six patients with homogeneous tumor blood supply (lower lacunarity) and higher rTBV were free of local recurrence during 24 months, while two of four patients with heterogeneous tumor blood supply (higher lacunarity) and lower rTBV died during follow-up due to tumor relapse. A weak correlation between FDG-PET and DCE-CT rTBV was observed (R(2)=0.1). CONCLUSION: FDG-PET/CT and DCT-CT are complementary methods for surveillance assessment in patients with HNT. Implementation of DCE-CT in the pretreatment FDG-PET/CT protocol may improve tumor outcome prediction.

On the effectiveness of ion range determination from in-beam PET data.

At present, in-beam positron emission tomography (PET) is the only method for in vivo and in situ range verification in ion therapy. At the GSI Helmholtzzentrum für Schwerionenforschung GmbH (GSI) Darmstadt, Germany, a unique in-beam PET installation has been operated from 1997 until the shut down of the carbon ion therapy facility in 2008. Therapeutic irradiation by means of (12)C ion beams of more than 400 patients have been monitored. In this paper a first quantitative study on the accuracy of the in-beam PET method to detect range deviations between planned and applied treatment in clinically relevant situations using simulations based on clinical data is presented. Patient treatment plans were used for performing simulations of positron emitter distributions. For each patient a range difference of + or - 6 mm in water was applied and compared to simulations without any changes. The comparisons were performed manually by six experienced evaluators for data of 81 patients. The number of patients required for the study was calculated using the outcome of a pilot study. The results indicate a sensitivity of (91 + or - 3)% and a specificity of (96 + or - 2)% for detecting an overrange, a reduced range is recognized with a sensitivity of (92 + or - 3)% and a specificity of (96 + or - 2)%. The positive and the negative predictive value of this method are 94% and 87%, respectively. The interobserver coefficient of variation is between 3 and 8%. The in-beam PET method demonstrated a high sensitivity and specificity for the detection of range deviations. As the range is a most indicative factor of deviations in the dose delivery, the promising results shown in this paper confirm the in-beam PET method as an appropriate tool for monitoring ion therapy.

Co-localisation of hypoxia and perfusion markers with parameters of glucose metabolism in human squamous cell carcinoma (hSCC) xenografts.

PURPOSE: To examine relationships between tumour hypoxia, perfusion and metabolic microenvironment at the microregional level in three different human squamous cell carcinomas (hSCC). MATERIALS AND METHODS: Nude mice bearing FaDu, UT-SCC-15, and UT-SCC-5 hSCC were injected with pimonidazole hypoxia and Hoechst perfusion markers. Bioluminescence imaging was used to determine spatial distribution of glucose and lactate content in serial tumour sections. Metabolite levels were grouped in 10 concentration ranges. Images were co-registered and at each concentration range the proportion of area stained for pimonidazole and Hoechst was determined in 11-13 tumours per tumour line. RESULTS: The spatial distribution of metabolites in pimonidazole hypoxic and Hoechst perfused areas is characterised by pronounced heterogeneity. In all three tumour lines glucose concentration decreased with increasing pimonidazole hypoxic fraction and increased with increasing perfused area at the microregional level. A weak albeit significant positive correlation between lactate concentration and pimonidazole hypoxic fraction was found only in UT-SCC-5. Lactate concentration consistently decreased with increasing perfused area in all three tumour lines. CONCLUSIONS: Both glucose consumption and supply may contribute to the microregional glucose levels. Microregional lactate accumulation in tumours may be governed by clearance potential. The extent of microregional hypoxia cannot be predicted from the lactate concentration indicating that both parameters need to be measured independently.

Direct time-of-flight for quantitative, real-time in-beam PET: a concept and feasibility study.

We extrapolate the impact of recent detector and scintillator developments, enabling sub-nanosecond coincidence timing resolution (tau), onto in-beam positron emission tomography (in-beam PET) for monitoring charged-hadron radiation therapy. For tau < or = 200 ps full width at half maximum, the information given by the time-of-flight (TOF) difference between the two opposing gamma-rays enables shift-variant, artefact-free in-beam tomographic imaging by means of limited-angle, dual-head detectors. We present the corresponding fast, TOF-based and backprojection-free, 3D reconstruction algorithm that, coupled with a real-time data acquisition and a fast detector encoding scheme, allows the sampled beta+-activity to be visualized in the object during the course of the irradiation. Despite the very low statistics scenario typical of in-beam PET, real-treatment simulations show that in-beam TOF-PET enables high-precision images to be obtained in real-time, either with closed-ring or with fixed, dual-head in-beam TOF-PET systems. The latter greatly alleviates the installation of in-beam PET at radiotherapeutic sites.

On the detector arrangement for in-beam PET for hadron therapy monitoring.

In-beam positron emission tomography (in-beam PET) is currently the only method for an in situ monitoring of highly tumour-conformed charged hadron therapy. At the experimental carbon ion tumour therapy facility, running at the Gesellschaft für Schwerionenforschung, Darmstadt, Germany, all treatments have been monitored by means of a specially adapted dual-head PET scanner. The positive clinical impact of this project triggered the construction of a hospital-based hadron therapy facility, with in-beam PET expected to monitor more delicate radiotherapeutic situations. Therefore, we have studied possible in-beam PET improvements by optimizing the arrangement of the gamma-ray detectors. For this, a fully 3D, rebinning-free, maximum likelihood expectation maximization algorithm applicable to several closed-ring or dual-head tomographs has been developed. The analysis of beta(+)-activity distributions simulated from real-treatment situations and detected with several detector arrangements allows us to conclude that a dual-head tomograph with narrow gaps yields in-beam PET images with sufficient quality for monitoring head and neck treatments. For monitoring larger irradiation fields, e.g. treatments in the pelvis region, a closed-ring tomograph was seen to be highly desirable. Finally, a study of the space availability for patient and bed, tomograph and beam portal proves the implementation of a closed-ring detector arrangement for in-beam PET to be feasible.