Deep multiple instance learning is the go-to in computational pathology as it efficiently handles weakly-supervised training. The use of attention mechanism allows to identify the tiles from the whole-slide images (WSIs) that are used to make a prediction, offering a first entry for interpretation. Recent studies proposed unsupervised approaches to summarize WSIs as a mixture of prototypical tiles, giving the opportunity to encode patients as a blend of different morphological patterns that can be translated to biological and clinical interpretation. These methods, albeit promising, fail to capture spatial context between tiles. In this paper, we propose to use spatial statistics to add spatial information to unsupervised patient encodings. More specifically, each tile is associated to its closest prototypical tile and Ripley's K function is used to characterize their spatial homogeneity across the slide. The evaluation of this new patient encoding on the Prostate cANcer graDe Assessment (PANDA) dataset demonstrates its superiority when compared to current patient encodings and MIL methods. Additionally, we show how to interpret this new patient encoding as a specific combination of morphological patterns and their spatial distribution at the WSI level.

Prediction of urinary toxicity after prostate cancer radiotherapy (RT) is remarkably challenging. Not only it is a multifaceted phenomenon, encompassing different symptoms (retention, dysuria, haematuria, etc.), but also a multifactorial problem, as it depends on both patient-specific clinical factors, individual biological parameters, and dosimetric patterns. Thus, there are a plethora of potential predictors compared to the paucity of available symptom-specific toxicity data. On top of that, in elder patients, urinary complications are not necessarily treatment-related which introduces important noise to the urinary toxicity assessment. In recent years, a growing interest in machine learning (ML) appears within the radiotherapy community. The goal of ML algorithms is to learn from existing data to recognize patterns in the population and make informed decisions. The purpose of this study was to implement two advanced heterogeneous ensemble methods, namely Stacking and Genetic Algorithm-based Weighted Average Voting for improving urinary toxicity prediction in the case of prostate cancer radiotherapy. Our analysis demonstrated that both GA-based Voting (AUC =0.66) and Stacking (AUC =0.80) outperformed the standard Weighted Voting classifier (AUC =0.66). In conclusion, Genetic Algorithm-based Weighted Average Voting may improve prediction performance compared to individual classifiers or conventional voting ensembles but at high computational cost. Stacking, on the other hand, appears significantly more powerful for predicting urinary toxicity at less computational cost.

Radiomics refers to the quantification of images by the extraction and analysis of a large number of features from different modalities, aiming to establish potential links between them and disease phenotypes. It can potentially predict the free-disease survival or allow the selection of patients at risk, thereby leading to the development of more personalized treatments. The development of robust prediction models is cumbersome as we deal with a high multidimensional problem, where a high number of features can be available but with a low number of individuals. To cope with this problem, we propose in this paper the use of Multiple Kernel Learning (MKL), which allows a selection of more relevant features and its optimal combination in a classification model. The method was evaluated on a dataset of patients of prostate cancer treated with radiotherapy, which is the second most prevalent cancer in men worldwide, for whom we predicted the risk of recurrence. MKL allowed the selection of 7 features out of 98 to build a reliable model with an accuracy of 94.7%, Sensitivity of 75%, and specificity of 97.78%. Compared to other classification methods, MKL achieved significantly higher performance, emerging like a suited methodology within radiomic studies.

In radiotherapy, computed tomography (CT) images are typically used for radiation treatment planning. Accurate segmentation of radiation sensitive healthy tissues, organs-at-risk (OARs), is important for radiation treatment planning for brain tumor. 3D Slicer has been applied in many medical applications including tumor segmentation on head MR images. However, to the best of our knowledge, there have been no studies using 3D Slicer for segmenting OARs on head CT images. This preliminary study evaluates the segmentation of seven OARs on head CTs using 3D Slicer. Results are comparable to state-of-the-art approaches but a larger dataset is required to verify the results.

Prostate cancer radiotherapy unavoidably involves the irradiation not only of the target volume, but also of healthy organs-at-risk, neighboring the prostate, likely causing adverse, toxicity-related side-effects. Specifically, in the case of urinary toxicity, these side effects might be associated with a variety of dosimetric, clinical and genetic factors, making its prediction particularly challenging. Given the inconsistency of available data concerning radiation-induced toxicity, it is crucial to develop robust models with superior predictive performance in order to perform tailored treatments. Machine Learning techniques emerge as appealing in this context, nevertheless without any consensus on the best algorithms to be used. This work proposes a comparison of several machine-learning strategies together with different minority class oversampling techniques for prediction of urinary toxicity following prostate cancer radiotherapy using dosimetric and clinical data. The performance of these classifiers was evaluated on the original dataset and using four different synthetic oversampling techniques. The area under the ROC curve (AUC) and the F-measure were employed to evaluate their performance. Results suggest that, regardless of the technique, oversampling always increases the prediction performance of the models (p=0.004). Overall, oversampling with Synthetic Minority Oversampling Technique (SMOTE) followed by Edited Nearest Neighbour algorithm (ENN) together with Regularized Discriminant Analysis (RDA) classifier provide the best performance (AUC=0.71).

External beam radiotherapy is extensively used to treat cervical carcinomas. A single planning CT scan enables the calculation of the dose distribution. The treatment is delivered over five weeks. Large per-treatment anatomical variations may hamper the dose delivery, with the potential of an organ-at-risk (OAR) overdose and a tumor underdose. To anticipate these deformations, a recent approach proposed three planning CTs with variable bladder volumes, which had the limitation of not covering all per-treatment anatomical variations. An original patient-specific population-based library has been proposed. It consisted of generating two representative anatomies, in addition to the standard planning CT anatomy. First, the cervix and bladder meshes of a population of 20 patients (314 images) were registered to an anatomical template, using a deformable mesh registration. An iterative point-matching algorithm was developed based on local shape context (histogram of polar or cylindrical coordinates and geodesic distance to the base) and on a topology constraint filter. Second, a standard principal component analysis (PCA) model of the cervix and bladder was generated to extract the dominant deformation modes. Finally, specific deformations were obtained using posterior PCA models, with a constraint representing the top of the uterus deformation. For a new patient, the cervix-uterus and bladder were registered to the template, and the patient's modeled planning library was built according to the model deformations. This method was applied following a leave-one-patient-out cross-validation. The performances of the modeled library were compared to those of the three-CT-based library, showing an improvement in both target coverage and OAR sparing.

External radiotherapy is a major clinical treatment for localized prostate cancer. Currently, computed tomography (CT) is used to delineate the prostate and to plan the radiotherapy treatment. However, CT images suffer from a poor soft-tissue contrast and do not allow an accurate organ delineation. On the contrary, magnetic resonance imaging (MRI) provides rich details and high soft-tissue contrast, allowing tumor detection. Thus, the intraindividual propagation of MRI delineations toward the planning CT may improve tumor targeting. In this paper, we introduce a new method to propagate MRI prostate delineations to the planning CT. In the first step, a random forest classification is performed to coarsely detect the prostate in the CT images, yielding a prostate probability membership for each voxel and a prostate hard segmentation. Then, the registration is performed using a new similarity metric which maximizes the probability and the collinearity between the normals of the manual registration (MR) existing contour and the contour resulting from the CT classification. The first study on synthetic data was performed to analyze the influence of the metric parameters with different levels of noise. Then, the method was also evaluated on real MR-CT data using manual alignments and intraprostatic fiducial markers and compared to a classically used mutual information (MI) approach. The proposed metric outperformed MI by 7% in terms of Dice score coefficient, by 3.14 mm the Hausdorff distance, and 2.13 mm the markers position errors. Finally, the impact of registration uncertainties on the treatment planning was evaluated, demonstrating the potential advantage of the proposed approach in a clinical setup to define a precise target.

The understanding of dose/side-effects relationships in prostate cancer radiotherapy is crucial to define appropriate individual's constraints for the therapy planning. Most of the existing methods to predict side-effects do not fully exploit the rich spatial information conveyed by the three-dimensional planned dose distributions. We propose a new classification method for three-dimensional individuals' doses, based on a new semi-nonnegative ICA algorithm to identify patients at risk of presenting rectal bleeding from a population treated for prostate cancer. The method first determines two bases of vectors from the population data: the two bases span vector subspaces, which characterize patients with and without rectal bleeding, respectively. The classification is then achieved by calculating the distance of a given patient to the two subspaces. The results, obtained on a cohort of 87 patients (at two year follow-up) treated with radiotherapy, showed high performance in terms of sensitivity and specificity.

This paper addresses the issue of cumulative dose estimation from cone beam computed tomography (CBCT) images in prostate cancer radiotherapy. It focuses on the dose received by the surfaces of the main organs at risk, namely the bladder and rectum. We have proposed both a surface-constrained dose accumulation approach and its extensive evaluation. Our approach relied on the nonrigid registration (NRR) of daily acquired CBCT images on the planning CT image. This proposed NRR method was based on a Demons-like algorithm, implemented in combination with mutual information metric. It allowed for different levels of geometrical constraints to be considered, ensuring a better point to point correspondence, especially when large deformations occurred, or in high dose gradient areas. The three following implementations: 1) full iconic NRR; 2) iconic NRR constrained with landmarks (LCNRR); 3) NRR constrained with full delineation of organs (DBNRR). To obtain reference data, we designed a numerical phantom based on finite-element modeling and image simulation. The methods were assessed on both the numerical phantom and real patient data in order to quantify uncertainties in terms of dose accumulation. The LCNRR method appeared to constitute a good compromise for dose monitoring in clinical practice.

We report in this communication a new formulation for the cost function of the well-known fuzzy C-means classification technique whereby we introduce weights. We derive the equations of this new weighted fuzzy C-means algorithm (WFCM) in the presence of additive and multiplicative bias field. We show that the weights can be designed in the same manner as prior probabilities commonly used in maximum a posteriori classifier (MAP) to introduce prior knowledge (e.g. using atlas), and increase robustness to noise (e.g. using Markov random field). Using prior probabilities of three popular MAP algorithms, we compare the performances of our proposed WFCM scheme using the simulated MRI T1W BrainWeb datasets, as well as five T1W MR patient scans. Our results show that WFCM achieves superior performances for low SNR conditions, whereas a Gaussian mixture model is desirable for high noise levels. WFCM allows rigorous comparison of fuzzy and probabilistic classifiers, and offers a framework where improvements can be shared between those two types of classifier.

This paper presents a new approach dealing with virtual exploratory navigation inside vascular structures. It is based on the notion of active vision in which only visual perception drives the motion of the virtual angioscope. The proposed fly-through approach does not require a premodeling of the volume dataset or an interactive control of the virtual sensor during the fly-through. Active navigation combines the on-line computation of the scene view and its analysis, to automatically define the three-dimensional sensor path. The navigation environment and the camera-like model are first sketched. The basic stages of the active navigation framework are then described: the virtual image computation (based on ray casting), the scene analysis process (using depth map), the navigation strategy, and the virtual path estimation. Experimental results obtained from phantom model and patient computed tomography data are finally reported.