Search

In this paper, we propose to quantitatively compare loss functions based on parameterized Tsallis-Havrda-Charvat entropy and classical Shannon entropy for the training of a deep network in the case of small datasets which are usually encountered in medical applications. Shannon cross-entropy is widely used as a loss function for most neural networks applied to the segmentation, classification and detection of images. Shannon entropy is a particular case of Tsallis-Havrda-Charvat entropy. In this work, we compare these two entropies through a medical application for predicting recurrence in patients with head-neck and lung cancers after treatment. Based on both CT images and patient information, a multitask deep neural network is proposed to perform a recurrence prediction task using cross-entropy as a loss function and an image reconstruction task. Tsallis-Havrda-Charvat cross-entropy is a parameterized cross entropy with the parameter $\alpha$. Shannon entropy is a particular case of Tsallis-Havrda-Charvat entropy for $\alpha$ = 1. The influence of this parameter on the final prediction results is studied. In this paper, the experiments are conducted on two datasets including in total 580 patients, of whom 434 suffered from head-neck cancers and 146 from lung cancers. The results show that Tsallis-Havrda-Charvat entropy can achieve better performance in terms of prediction accuracy with some values of $\alpha$., Comment: 11 pages, 3 figures

Background Segmentation is a crucial step in treatment planning that directly impacts dose distribution and optimization. The aim of this study was to evaluate the inter-individual variability of common cranial organs at risk (OAR) delineation in neurooncology practice. Methods Anonymized simulation contrast-enhanced CT and MR scans of one patient with a solitary brain metastasis was used for delineation and analysis. Expert professionals from 16 radiotherapy centers involved in brain structures delineation were asked to segment 9 OAR on their own treatment planning system. As reference, two experts in neurooncology, produced a unique consensual contour set according to guidelines. Overlap ratio, Kappa index (KI), volumetric ratio, Commonly Contoured Volume, Supplementary Contoured Volume were evaluated using Artiview™ v 2.8.2—according to occupation, seniority and level of expertise of all participants. Results For the most frequently delineated and largest OAR, the mean KI are often good (0.8 for the parotid and the brainstem); however, for the smaller OAR, KI degrade (0.3 for the optic chiasm, 0.5% for the cochlea), with a significant discrimination (p Association des Neuro-Oncologue d’Expression Française society performed better in all indicators compared to non-members (p p Conclusion Our study illustrates the heterogeneity in normal structures contouring between professionals. We emphasize the need for cerebral OAR delineation harmonization—that is a major determinant of therapeutic ratio and clinical trials evaluation.

International audience; Radiomics is defined as the extraction of a large quantity of quantitative image features. The different radiomic indexes that have been proposed in the literature are described as well as the various factors that have an impact on the robustness of these indexes. We will see that several hundred quantitative features can be extracted per lesion and imaging modality. The ever-growing number of features studied raises the question of the statistical method of analysis used.This review addresses the research supporting the clinical use of radiomics in oncology in the staging of disease, discrimination between healthy and pathological tissues, the identification of genetic features, the prediction of patient survival, the response to treatment, the recurrence after radiotherapy and chemoradiotherapy and the side effects.Based on the existing literature, it remains difficult to identify features that should be used for current clinical practice.

[EN] The human cerebellum plays an essential role in motor control, is involved in cognitive function (i.e., attention, working memory, and language), and helps to regulate emotional responses. Quantitative in-vivo assessment of the cerebellum is important in the study of several neurological diseases including cerebellar ataxia, autism, and schizophrenia. Different structural subdivisions of the cerebellum have been shown to correlate with differing pathologies. To further understand these pathologies, it is helpful to automatically parcellate the cerebellum at the highest fidelity possible. In this paper, we coordinated with colleagues around the world to evaluate automated cerebellum parcellation algorithms on two clinical cohorts showing that the cerebellum can be parcellated to a high accuracy by newer methods. We characterize these various methods at four hierarchical levels: coarse (i.e., whole cerebellum and gross structures), lobe, subdivisions of the vermis, and the lobules. Due to the number of labels, the hierarchy of labels, the number of algorithms, and the two cohorts, we have restricted our analyses to the Dice measure of overlap. Under these conditions, machine learning based methods provide a collection of strategies that are efficient and deliver parcellations of a high standard across both cohorts, surpassing previous work in the area. In conjunction with the rank-sum computation, we identified an overall winning method., The data collection and labeling of the cerebellum was supported in part by the NIH/NINDS grant R01 NS056307 (PI: J.L. Prince) and NIH/NIMH grants R01 MH078160 & R01 MH085328 (PI: S.H. Mostofsky). PMT is supported in part by the NIH/NIBIB grant U54 EB020403. CERES2 development was supported by grant UPV2016-0099 from the Universitat Politecnica de Valencia (PI: J.V. Manjon); the French National Research Agency through the Investments for the future Program IdEx Bordeaux (ANR-10-IDEX-03-02, HL-MRI Project; PI: P. Coupe) and Cluster of excellence CPU and TRAIL (HR-DTI ANR-10-LABX-57; PI: P. Coupe). Support for the development of LiviaNET was provided by the National Science and Engineering Research Council of Canada (NSERC), discovery grant program, and by the ETS Research Chair on Artificial Intelligence in Medical Imaging. The authors wish to acknowledge the invaluable contributions offered by Dr. George Fein (Dept. of Medicine and Psychology, University of Hawaii) in preparing this manuscript.

International audience; Purpose: Radiotherapy aims at delivering the highest possible dose to the tumor while minimizing the irradiation of surrounding healthy tissue, and especially to the organs at risk (OARs). Therefore, accurate delineation of 15 OARs is required for radiation treatment planning (RTP). In thoracic oncology, delineation of some OARs remains manual, making the task time consuming and prone to inter observer variability. Various (semi-) automatic approaches have been proposed to segment OARs on CT but the task still remains challenging. Here, a system to interactively segment OARs in thoracic oncology on CT images is presented and its clinical acceptability evaluated. 20 Methods: The proposed framework has been implemented using MITK platform. User interaction lies in the easy definition of few manual seeds for the OARs and background using a 'paintbrush' tool, which can be interactively added in any view (axial, sagittal or coronal), and is subsequently propagated within the whole volume. Once the user is content with the seeds placement, the system automatically performs the segmentation. If the outcome is not satisfying, the user can modify the seeds, which involves adding and/or removing existing seeds, and perform 25 again the automatic segmentation. Number of tries has been limited to five in the current study. If after the five modifications the segmentation result is not sufficient to be usable in the RTP, the user shall reject it; otherwise, he shall accept it. A hybrid approach based on our previous work [1, 2] which combines watershed transformation and graph cuts [3] is used for the segmentation task. Results: The system was evaluated on multivendor CT datasets of 10 patients presenting from early stage to 30 locally advanced NSCLC or pulmonary metastases. OARs taken into consideration in this study were: heart, lungs, oesophagus, proximal bronchus tree, spinal canal and trachea. Interactive contours were generated by a physician using the proposed system. Delineation of the OARs obtained with the presented system was approved to be usable for RTP in more than 90% of the cases, excluding the oesophagus, which segmentation was never approved (Fig 1). On the accepted reported cases, more than 90% of the interactive contours reached a Dice 35 Similarity Coefficient higher than 0.7 with respect to manual segmentations (Fig 2). Therefore, our interactive delineation approach allows users to generate contours of sufficient quality to be used in RTP up to three times faster than manually. Conclusions: An interactive, accurate and easy-to-use computer-assisted system for OARs segmentation in tho-racic oncology was presented and clinically evaluated. The introduction of the proposed approach in clinical 40 routine might offer a valuable new option to radiation therapists (RTTs) in performing OARs delineation task. Consequently, further experiments have been conducted on larger databases and with the participation of additional RTTs to investigate its potential use in daily clinical practice [4].

International audience; Purpose: Despite post-operative radiotherapy (RT) of glioblastoma (GBM), local tumor regrowth occur in irradiated areas and are responsible for poor outcome. Identification of sites with high probability of recurrence is a promising way to define new target volumes for dose escalation in RT treatments. This study aims at assessing the value of multi-parametric magnetic resonance (mp-MR) data acquired before RT treatment in the identification of regions at risk of relapse. Methods: Ten newly diagnosed GBM patients included in a clinical trial, treated in the reference arm of 60 Gy and Temozolomide, underwent magnetic resonance imaging (MRI) and MR spectroscopy (MRSI) before RT treatment and every 2 months until relapse. Quantification of MRSI was improved following the spatio-spectral regularization approach proposed in [1]. The site of relapse was considered as the new appearing contrast-enhancing (CE) areas on T1-weighted images after gadolinium injection (T1-Gd). Using a set of mp-MR data acquired before RT treatment as input, a supervised learning system was trained to generate a probability map of CE appearance of GBM. Since it has recently shown a great performance in some other brain-related classification problems, such as segmentation of brain tumor [2] or some brain structures [3], support vector machines (SVM) [4] was chosen as the learning technique. To fed the classifier, T1-Gd and FLAIR image intensities, Choline-over-NAA, Choline-over-Creatine and Lac-over-NAA metabolite ratios, and metabolite heights were used. The resolution of the MRI images was lowered to the one of the MRSI grid by averaging MRI pixel intensities within each MRSI voxel (400 MRSI voxels were considered for each subject). The region of CE was manually contoured on both the pre-RT and post-RT T1-Gd images by experienced medical staff. All voxels that enhanced at the pre-RT exam were excluded from further consideration. The learning system was trained and tested using leave-one-out-cross-validation (LOOCV) with all the patients. A grid-search strategy was employed for parameter optimization. Results: For comparison purposes, generated probability maps were thresholded with a value of 0.5. Thus, only voxels with values higher than 0.5 on the probability map were considered as relapse. The sensitivity and specificity of the proposed system were 0.80 (0.19) and 0.87 (0.09), respectively. For our data, standard Choline-to-NAA index (CNI) achieved a sensitivity of 0.62 (0.25) and a specificity of 0.63 (0.13). In addition, the receiver operating characteristic (ROC) curve also shows that the presented approach outperforms CNI (Fig 1.). The SVM-based results had lower variation across patients than CNI. An example of a probability map of relapse areas generated by the proposed approach is shown in Fig.2. Relapse regions predicted by the learning scheme are in high accordance with the manually contoured region. Conclusions: A learning system based on SVM trained with mp-MR data has been presented. Reported results imply that this learning scheme can provide a probability map of the area of relapse of GBM in a stable and accurate manner, as shown in Fig.1. This study suggests the potential of multi-parametric MR data in addressing specific questions in GBM imaging.

International audience; AbstractPURPOSE: Accurate delineation of organs at risk (OARs) on computed tomography (CT) image is required for radiation treatment planning (RTP). Manual delineation of OARs being time consuming and prone to high interobserver variability, many (semi-) automatic methods have been proposed. However, most of them are specific to a particular OAR. Here, an interactive computer-assisted system able to segment various OARs required for thoracic radiation therapy is introduced.METHODS: Segmentation information (foreground and background seeds) is interactively added by the user in any of the three main orthogonal views of the CT volume and is subsequently propagated within the whole volume. The proposed method is based on the combination of watershed transformation and graph-cuts algorithm, which is used as a powerful optimization technique to minimize the energy function. The OARs considered for thoracic radiation therapy are the lungs, spinal cord, trachea, proximal bronchus tree, heart, and esophagus. The method was evaluated on multivendor CT datasets of 30 patients. Two radiation oncologists participated in the study and manual delineations from the original RTP were used as ground truth for evaluation.RESULTS: Delineation of the OARs obtained with the minimally interactive approach was approved to be usable for RTP in nearly 90% of the cases, excluding the esophagus, which segmentation was mostly rejected, thus leading to a gain of time ranging from 50% to 80% in RTP. Considering exclusively accepted cases, overall OARs, a Dice similarity coefficient higher than 0.7 and a Hausdorff distance below 10 mm with respect to the ground truth were achieved. In addition, the interobserver analysis did not highlight any statistically significant difference, at the exception of the segmentation of the heart, in terms of Hausdorff distance and volume difference.CONCLUSIONS: An interactive, accurate, fast, and easy-to-use computer-assisted system able to segment various OARs required for thoracic radiation therapy has been presented and clinically evaluated. The introduction of the proposed system in clinical routine may offer valuable new option to radiation oncologists in performing RTP.

International audience; Delineation of organs at risk (OARs) is a crucial step in surgical and treatment planning in brain cancer, where precise OARs volume delineation is required. However, this task is still often manually performed, which is time-consuming and prone to observer variability. To tackle these issues a deep learning approach based on stacking denoising auto-encoders has been proposed to segment the brainstem on magnetic resonance images in brain cancer context. Additionally to classical features used in machine learning to segment brain structures, two new features are suggested. Four experts participated in this study by segmenting the brainstem on 9 patients who underwent radiosurgery. Analysis of variance on shape and volume similarity metrics indicated that there were significant differences (p

International audience; Abstract:Purpose: This study aimed to evaluate the impact of consensus algorithms on segmentation results when applied on clinical PET images. In particular, how majority vote or STAPLE algorithms could improve the final result in terms of accuracy and reproducibility when combining three semi-automatic segmentation algorithms.Methods: Three published approaches of segmentation (contrast-oriented, possibility theory and adaptive thresholding) and two consensus algorithms, majority vote and STAPLE, were implemented in a single software platform (Artiview®). Four clinical datasets including different locations (thorax, breast, abdomen) or pathologies (NSCLC primary tumours, metastasis, lymphoma) were used to evaluate accuracy and reproducibility of the consensus approach in comparison with pathology ground truth or CT – ground truth surrogate.Results: Our results reflect the variable performance of individual segmentation algorithms for lesions of different tumour entities that is for PET images that differ in resolution, contrast and image noise. Independent on location and pathology of the lesion, however, the consensus method displays improved volume segmentation accuracy compared to the worst performing individual method in the majority of cases and is close to the best performing method in many cases. In addition, the implementation reveals high reproducibility of the segmentation results against small changes in the respective starting conditions. No significant difference between STAPLE and majority vote algorithms was found.Conclusion: This study shows that combining different PET-segmentation methods by application of a consensus algorithm offers robustness against the variable performance of individual segmentation methods and is therefore useful for radiation oncology purposes. It might also be relevant for other scenarios like the joining of expert recommendations in clinical routine and trials or the generation of multi-observer generated contours for standardisation of automatic contouring.

The aim of this study was to evaluate the impact of consensus algorithms on segmentation results when applied to clinical PET images. In particular, whether the use of the majority vote or STAPLE algorithm could improve the accuracy and reproducibility of the segmentation provided by the combination of three semiautomatic segmentation algorithms was investigated. Three published segmentation methods (contrast-oriented, possibility theory and adaptive thresholding) and two consensus algorithms (majority vote and STAPLE) were implemented in a single software platform (Artiview®). Four clinical datasets including different locations (thorax, breast, abdomen) or pathologies (primary NSCLC tumours, metastasis, lymphoma) were used to evaluate accuracy and reproducibility of the consensus approach in comparison with pathology as the ground truth or CT as a ground truth surrogate. Variability in the performance of the individual segmentation algorithms for lesions of different tumour entities reflected the variability in PET images in terms of resolution, contrast and noise. Independent of location and pathology of the lesion, however, the consensus method resulted in improved accuracy in volume segmentation compared with the worst-performing individual method in the majority of cases and was close to the best-performing method in many cases. In addition, the implementation revealed high reproducibility in the segmentation results with small changes in the respective starting conditions. There were no significant differences in the results with the STAPLE algorithm and the majority vote algorithm. This study showed that combining different PET segmentation methods by the use of a consensus algorithm offers robustness against the variable performance of individual segmentation methods and this approach would therefore be useful in radiation oncology. It might also be relevant for other scenarios such as the merging of expert recommendations in clinical routine and trials or the multiobserver generation of contours for standardization of automatic contouring.

International audience; Delineating critical structures of the brain is required for advanced radiotherapy technologies to determine whether the dose from the proposed treatment will impair the functional-ity of those structures. Employing an automatic segmentation computer module in the radiation oncology treatment planning process has the potential to significantly increase the efficiency , cost-effectiveness, and, ultimately, clinical outcome of patients undergoing radiation therapy. Atlas-based seg-mentation has shown to be a suitable tool for the segmentation of large structures such as the brainstem or the cerebellum. However, smaller structures such as the optic nerves are more difficult to segment. In this work, we present a novel approach to automatically segment the optic nerves, which is based on Support Vector Machines (SVM). Compared to state of the art methods, the presented method obtained a better performance in regards to accuracy, robustness and processing time, being a suitable trade-off between these three factors.

International audience; AbstractThis work covers the current state of the art with regard to approaches to segment subcortical brain structures. A huge range of diverse methods have been presented in the literature during the last decade to segment not only one or a constrained number of structures, but also a complete set of these subcortical regions. Special attention has been paid to atlas based segmentation methods, statistical models and deformable models for this purpose. More recently, the introduction of machine learning techniques, such as artificial neural networks or support vector machines, has helped the researchers to optimize the classification problem. These methods are presented in this work, and their advantages and drawbacks are further discussed. Although these methods have proved to perform well, their use is often limited to those situations where either there are no lesions in the brain or the presence of lesions does not highly vary the brain anatomy. Consequently, the development of segmentation algorithms that can deal with such lesions in the brain and still provide a good performance when segmenting subcortical structures is highly required in practice by some clinical applications, such as radiotherapy or radiosurgery.

Alexandre Huat, Sébastien Thureau, David Pasquier, Isabelle Gardin, Romain Modzelewski, David Gibon, Juliette Thariat and Vincent Grégoire were not included as authors in the original publication [...].