Search

Purpose Total marrow (and lymphoid) irradiation (TMI-TMLI) is limited by the couch travel range of modern linacs, which forces the treatment delivery to be split into two plans with opposite orientations: a head-first supine upper-body plan, and a feet-first supine lower extremities plan. A specific field junction is thus needed to obtain adequate target coverage in the overlap region of the two plans. In this study, an automatic procedure was developed for field junction creation and lower extremities plan optimization. Methods Ten patients treated with TMI-TMLI at our institution were selected retrospectively. The planning of the lower extremities was performed automatically. Target volume parameters (CTV_J‑V98% > 98%) at the junction region and several dose statistics (D98%, Dmean, and D2%) were compared between automatic and manual plans. The modulation complexity score (MCS) was used to assess plan complexity. Results The automatic procedure required 60–90 min, depending on the case. All automatic plans achieved clinically acceptable dosimetric results (CTV_J‑V98% > 98%), with significant differences found at the junction region, where Dmean and D2% increased on average by 2.4% (p p Conclusion The developed procedure allowed treatment planning of TMI-TMLI to be streamlined, increasing efficiency and standardization, preventing human errors, while maintaining the dosimetric plan quality and complexity of manual plans. Automated strategies can simplify the future adoption and clinical implementation of TMI-TMLI treatments in new centers.

Purpose To identify the best transfer learning approach in the identification of the most frequent abnormalities on chest radiograph (CXR) using embeddings extracted from pre-trained convolutional neural networks (CNNs). Explainable AI (XAI) model was applied to interpret black-box models’ predictions and assess its performance. Methods and Materials: Seven CNNs were trained on CheXpert. Three different transfer learning approaches were thereafter applied to a local dataset. Ensembling of the classification results was performed using simple and entropy-weighted averaging. Model performance was assessed by using the area under the curve (AUC). We applied Grad-CAM (an XAI model) to produce a saliency map, highlighting areas of the image relevant to the classification of the output class, both on single images and on a 200 CXRs sample. Grad-CAM maps were compared to regions of interest extracted manually. The training time was recorded. Results The best transfer learning model was the one that used image embeddings and Random Forest with simple averaging, whit an average AUC of 0.856. Grad-CAM maps showed that the models were focusing on specific features of each CXR. The training time was less than five minutes. Conclusions CNNs pre-trained on a large public dataset of medical images can be exploited as feature extractors for a task of interest. The extracted image embeddings contain relevant information to train an additional classifier with good performance on an independent dataset demonstrating to be the optimal transfer learning strategy, overcoming the need for large private datasets, high computational resources, and long training times.

Background: The total marrow and lymph node irradiation (TMLI) target includes the bones, spleen, and lymph node chains, with the latter being the most challenging structures to contour. We evaluated the impact of introducing internal contour guidelines to reduce the inter- and intraobserver lymph node delineation variability in TMLI treatments. Methods: A total of 10 patients were randomly selected from our database of 104 TMLI patients so as to evaluate the guidelines’ efficacy. The lymph node clinical target volume (CTV_LN) was recontoured according to the guidelines (CTV_LN_GL_RO1) and compared to the historical guidelines (CTV_LN_Old). Both topological (i.e., Dice similarity coefficient (DSC)) and dosimetric (i.e., V95 (the volume receiving 95% of the prescription dose) metrics were calculated for all paired contours. Results: The mean DSCs were 0.82 ± 0.09, 0.97 ± 0.01, and 0.98 ± 0.02, respectively, for CTV_LN_Old vs. CTV_LN_GL_RO1, and between the inter- and intraobserver contours following the guidelines. Correspondingly, the mean CTV_LN-V95 dose differences were 4.8 ± 4.7%, 0.03 ± 0.5%, and 0.1 ± 0.1%. Conclusions: The guidelines reduced the CTV_LN contour variability. The high target coverage agreement revealed that historical CTV-to-planning-target-volume margins were safe, even if a relatively low DSC was observed.

Chest X-Ray (CXR) is one of the most common diagnostic techniques used in everyday clinical practice all around the world. We hereby present a work which intends to investigate and analyse the use of Deep Learning (DL) techniques to extract information from such images and allow to classify them, trying to keep our methodology as general as possible and possibly also usable in a real world scenario without much effort, in the future. To move in this direction, we trained several beta-Variational Autoencoder (beta-VAE) models on the CheXpert dataset, one of the largest publicly available collection of labeled CXR images; from these models, latent features have been extracted and used to train other Machine Learning models, able to classify the original images from the features extracted by the beta-VAE. Lastly, tree-based models have been combined together in ensemblings to improve the results without the necessity of further training or models engineering. Expecting some drop in pure performance with the respect to state of the art classification specific models, we obtained encouraging results, which show the viability of our approach and the usability of the high level features extracted by the autoencoders for classification tasks., Comment: 8 pages, 5 figures

Chest X-ray (CXR) is perhaps the most frequently-performed radiological investigation globally. In this work, we present and study several machine learning approaches to develop automated CXR diagnostic models. In particular, we trained several Convolutional Neural Networks (CNN) on the CheXpert dataset, a large collection of more than 200k CXR labeled images. Then, we used the trained CNNs to compute embeddings of the CXR images, in order to train two sets of tree-based classifiers from them. Finally, we described and compared three ensembling strategies to combine together the classifiers trained. Rather than expecting some performance-wise benefits, our goal in this work is showing that the above two methodologies, i.e., the extraction of image embeddings and models ensembling, can be effective and viable to solve tasks that require medical imaging understanding. Our results in that perspective are encouraging and worthy of further investigation., Comment: Accepted at IJCNN 2021

Purpose The present scoping review aims to assess the non-inferiority of distributed learning over centrally and locally trained machine learning (ML) models in medical applications. Methods We performed a literature search using the term “distributed learning” OR “federated learning” in the PubMed/MEDLINE and EMBASE databases. No start date limit was used, and the search was extended until July 21, 2020. We excluded articles outside the field of interest; guidelines or expert opinion, review articles and meta-analyses, editorials, letters or commentaries, and conference abstracts; articles not in the English language; and studies not using medical data. Selected studies were classified and analysed according to their aim(s). Results We included 26 papers aimed at predicting one or more outcomes: namely risk, diagnosis, prognosis, and treatment side effect/adverse drug reaction. Distributed learning was compared to centralized or localized training in 21/26 and 14/26 selected papers, respectively. Regardless of the aim, the type of input, the method, and the classifier, distributed learning performed close to centralized training, but two experiments focused on diagnosis. In all but 2 cases, distributed learning outperformed locally trained models. Conclusion Distributed learning resulted in a reliable strategy for model development; indeed, it performed equally to models trained on centralized datasets. Sensitive data can get preserved since they are not shared for model development. Distributed learning constitutes a promising solution for ML-based research and practice since large, diverse datasets are crucial for success.

We present an educational application of virtual reality that we created to help students gain an in-depth understanding of the internal structure of crystals and related key concepts. Teachers can use it to give lectures to small groups (10-15) of students in a shared virtual environment, both remotely (with teacher and students in different locations) and locally (while sharing the same physical space). Lectures can be recorded, stored in an online repository, and shared with students who can either review a recorded lecture in the same virtual environment or can use the application for self-studying by exploring a large collection of available crystal structures. We validated our application with human subjects receiving positive feedback.The online version contains supplementary material available at 10.1007/s11042-022-13410-0https://doi.org/10.1007/s11042-022-13410-0.

Space debris represents a threat to space missions and operational satellites. Failing to control its growth might lead to the inability to use near-Earth space. However, this issue is still largely unknown to most people. In this paper, we present an educational experience in virtual reality created to raise awareness about the problem of space debris. The application exploits the entity-component-system (ECS) programming pattern to manage around 20000 orbiting objects with a high frame rate to convey a fluid experience. We preliminarily validated our application, in terms of usability as well as quality of user experience, during several events involving both a broad audience (e.g., citizens of all ages, from teenagers to elders) and an experienced audience (e.g., engineering students enrolled in the aerospace degree). The results of the evaluation were extremely positive, showing once again that virtual reality can be an effective means to engage people in captivating and interactive activities, making them experience what can only be imagined - the thousands pieces of space debris surrounding our planet.

Deep Learning is a promising approach to either automate or simplify several tasks in the healthcare domain. In this work, we introduce SegAN-CAT, an end-to-end approach to brain tumor segmentation in Magnetic Resonance Images (MRI), based on Adversarial Networks. In particular, we extend SegAN, successfully applied to the same task in a previous work, in two respects: (i) we used a different model input and (ii) we employed a modified loss function to train the model. We tested our approach on two large datasets, made available by the Brain Tumor Image Segmentation Benchmark (BraTS). First, we trained and tested some segmentation models assuming the availability of all the major MRI contrast modalities, i.e., T1-weighted, T1 weighted contrast enhanced, T2-weighted, and T2-FLAIR. However, as these four modalities are not always all available for each patient, we also trained and tested four segmentation models that take as input MRIs acquired with a single contrast modality. Finally, we proposed to apply transfer learning across different contrast modalities to improve the performance of these single-modality models. Our results are promising and show that not only SegAN-CAT is able to outperform SegAN when all the four modalities are available, but also that transfer learning can actually lead to better performances when only a single modality is available.

We applied Generative Adversarial Networks (GANs) to learn a model of DOOM levels from human-designed content. Initially, we analyzed the levels and extracted several topological features. Then, for each level, we extracted a set of images identifying the occupied area, the height map, the walls, and the position of game objects. We trained two GANs: one using plain level images, one using both the images and some of the features extracted during the preliminary analysis. We used the two networks to generate new levels and compared the results to assess whether the network trained using also the topological features could generate levels more similar to human-designed ones. Our results show that GANs can capture intrinsic structure of DOOM levels and appears to be a promising approach to level generation in first person shooter games.

Tile coding is an effective reinforcement learning method that uses a rather ingenious generalization mechanism based on (1) a carefully designed parameter setting and (2) the assumption that nearby states in the problem space will correspond to similar payoff predictions in the action-value function. Previously, we extended XCSF with tile coding prediction and compared it to tabular tile coding, showing that (1) XCSF performs as well as parameter-optimized tile coding, while also (2) evolving individualized parameter settings in each problem subspace. Our comparison was based on a set of well-known reinforcement learning environments (2D Gridworld and the Mountain Car) that involved no action-value discontinuities and so posed no challenge to tabular tile coding. In this paper, we go a step further and test XCSF with tile coding on a set of problems designed to challenge tile coding by introducing discontinuities in the action value landscape. The new testbed (called MazeWorld) extends 2D Gridworld with impenetrable obstacles, a conceptually simple modification that can dramatically increase the problem complexity for tabular tile coding. We compare four versions of XCSF with tile coding (each adapting a different set of parameters) to tabular tile coding on four problems of increasing complexity. We show that our system (1) needs fewer training problems than standard tile coding to reach an optimal policy; (2) it can evolve adequate coding parameters in each subspace without any previous knowledge; and that (3) even when XCSF is not allowed to evolve these parameters, the genetic algorithm will still adapt classifier conditions to properly decompose the problem into subspaces thus being much less sensitive to the parameter settings than tabular tile coding.

The design of video game levels is a complex and critical task. Levels need to elicit fun and challenge while avoiding frustration at all costs. In this paper, we present a framework to assist designers in the creation of levels for 2D platformers. Our framework provides designers with a toolbox (i) to create 2D platformer levels, (ii) to estimate the difficulty and probability of success of single jump actions (the main mechanics of platformer games), and (iii) a set of metrics to evaluate the difficulty and probability of completion of entire levels. At the end, we present the results of a set of experiments we carried out with human players to validate the metrics included in our framework., Submitted to the IEEE Game Entertainment and Media Conference 2018

In car racing, blocking refers to maneuvers that can prevent, disturb or completely block an overtaking action by an incoming car. In this paper, we present an advanced overtaking behavior that is able to deal with opponents implementing blocking strategies of various difficulty level. The behavior we developed has been integrated in an existing fuzzy-based architecture for driving simulated cars and tested using The Open Car Racing Simulator (TORCS). We compared a driver implementing our overtaking strategy against four of the bots available in the TORCS distribution and Simplix, a state-of-the-art driver which won several competitions. The comparison was carried out against opponents implementing three blocking strategies of increasing difficulty and two different scenarios: (i) a basic scenario with one opponent on a straight stretch to overtake as quickly as possible; (ii) an advanced scenario involving a race on a non-trivial track against several opponents. The results from the basic scenario show that our strategy can always overtake the opponent car; in particular, our strategy is slightly more risky than the other ones and may result in a little damage, however, all the other controllers show a more careful and safe policy that often prevents them to complete an overtaking maneuver. When racing against several opponents on complex tracks, our strategy results in the best trade-off between the time spent being blocked by an opponent ahead and the number of overtaking maneuvers completed.

The fight-or-flight response is a typical physiological reaction to an imminent threat, which prepares the body to fight or flee. Such a mechanism can be also exploited to some extent in game design to elicit strong emotions in players. However, in first person shooters, fleeing is often a not viable option, making fight inevitable. In this work, we suggest that map design can be effectively used to deal with this issue. In particular, we presented a procedural content generation approach, based on evolutionary computation, to evolve maps for Cube 2, an open source first person shooter. Our results show that the design of evolved maps is effectively able to foster the emergence of a fleeing behavior, even in AI controlled characters specifically designed to fight and chase the opponents.

Every year the Genetic and Evolutionary Computation Conference (GECCO) offers a diverse program of competitions that represents a wonderful opportunity to address challenging problems while competing with other members of the evolutionary computation (EC) community. These competitions also showcase riveting applications of EC in the top conference of the field. In 2013 GECCO offered six different competitions, which received a total of 31 submissions

The paper considers the evolution of a particular class of networks of identical chaotic oscillators, namely that of ecological networks. In these networks, nodes represent patches where a certain number of plant and animal populations interact on ecological timescale, arcs represent migration flows due to dispersal, and Darwinian evolution is responsible for variations, on a longer evolutionary timescale, of the demographic parameters characterizing the populations. Up to now, this problem has been mainly studied with reference to single-population patches described by one-dimensional discrete-time models and by considering only the dispersal rates of migrating populations as an evolving trait. Here, we propose a method of investigation which allows to study multipopulation patches described by continuous-time models with evolving traits influencing various demographic parameters (including or not dispersal). The method is casted within the frame of the so-called master stability function approach for the analysis of synchronization of coupled systems, and the results obtained in a first and very simple application support the conjecture that evolution drives ecological networks toward weak forms of synchronization.

Design a set of weapons for a competitive first person shooter (FPS) is not an easy task: it involves several challenges, such as balancing, diversity, and novelty. In this paper, we combine procedural content generation and evolutionary computation to solve this complex task. In particular, we introduce a novel approach to procedurally generate weapons for Unreal Tournament III, a popular commercial FPS. In our approach, each weapon is represented as a parameters vector and evaluated based on game statistics. We tested our approach on three different application scenarios that represents three typical design problems: (i) create a balanced set of weapons, (ii) improve a given weapon with the least possible changes, and (iii) create a set of weapons with some specific design goals. Finally, we also performed a preliminary user study to validate our work. Our results are promising and suggest that the proposed approach might be successfully used to support the weapon design process.

In this work, we introduce Volcano, a tool for the procedural generation of 3D models of swords. Unlike common procedural content generation tools, it exploits interactive evolution to reduce as much as possible the effort of the users during the generation process. Indeed, Volcano allows to forge the desired type of swords through a rather simple visual exploration of the design space. The 3D models generated with the tool can be directly used as game assets or further developed with a standard modeling software. A prototype of Volcano was tested by 30 users, including both students and game developers. The feedbacks received are very positive: tools like Volcano might be useful both for players, to create user contents, and for developers, to speed-up the design of game contents.

TrackGen is an online tool for the generation of tracks for two open-source 3D car racing games (TORCS and Speed Dreams). It integrates interactive evolution with procedural content generation and comprises two components: (i) a web frontend that maintains the database of all the evolved populations and manages the interaction with users (by collecting users evaluations and providing access to the evolved tracks) and (ii) an evolutionary/content-generation backend that runs both the evolutionary algorithm and generates the actual game content that is available through the web frontend. The first prototype of the tool was presented in July 2011 but advertised only to researchers; the first official version which generated tracks only for TORCS was released to the game community in September 2011; due to the many requests, we released a new version soon afterwards, in January 2012, with support for Speed Dreams (the fork of TORCS focused on visual realism and graphic quality) that has been online since then. From January 2012 until July 2014, TrackGen had more than 7600 unique visitors who visited the website around 11,500 times and viewed 85,500 pages; it was employed to evolve more than 8853 tracks, and it was used to download 12,218 tracks. Some of the tracks evolve by our system have been also included in the TORCS distribution.

In this paper, we investigate the application of evolutionary computation to the automatic generation of tracks for high-end racing games. The idea underlying our approach is that diversity is a major source of challenge/interest for racing tracks and, eventually, might play a key role in contributing to the player's fun. In particular, we focus on the diversity of a track in terms of its shape (i.e., the number and the assortment of turns and straights it contains), and in terms of driving experience it provides (i.e., the range of speeds achievable while driving on the track). We define two fitness functions that capture our idea of diversity as the entropy of the track's curvature and speed profiles. We apply both a single-objective and a multiobjective real-coded genetic algorithm (GA) to evolve tracks involving both a wide variety of turns and straights and also a large range of driving speeds. The results we report show that both single-objective and multiobjective approaches can successfully evolve tracks with a high degree of diversity both in terms of shape and achievable speeds.

High-Level Synthesis (HLS) is the process of developing digital circuits from behavioral specifications. It involves three interdependent and NP-complete optimization problems: (i) the operation scheduling, (ii) the resource allocation, and (iii) the controller synthesis. Evolutionary Algorithms have been already effectively applied to HLS to find good solution in presence of conflicting design objectives. In this paper, we present an evolutionary approach to HLS that extends previous works in three respects: (i) we exploit the NSGA-II, a multi-objective genetic algorithm, to fully automate the design space explorationwithout the need of any human intervention, (ii) we replace the expensive evaluation process of candidate solutions with a quite accurate regression model, and (iii) we reduce the number of evaluations with a fitness inheritance scheme. We tested our approach on several benchmark problems. Our results suggest that all the enhancements introduced improve the overall performance of the evolutionary search.

XCSF is the extension of XCS in which classifier prediction is computed as a linear combination of classifier inputs and a weight vector associated to each classifier. XCSF can exploit classifiers' computable prediction to evolve accurate piecewise linear approximations of functions. In this paper, we take XCSF one step further and show how XCSF can be easily extended to allow polynomial approximations. We test the extended version of XCSF on various approximation problems and show that quadratic/cubic approximations can be used to significantly improve XCSF's generalization capabilities.

We apply XCS with computed prediction (XCSF) to tackle multistep reinforcement learning problems involving continuous inputs. In essence we use XCSF as a method of generalized reinforcement learning. We show that in domains involving continuous inputs and delayed rewards XCSF can evolve compact populations of accurate maximally general classifiers which represent the optimal solution to the target problem. We compare the performance of XCSF with that of tabular Q-learning adapted to the continuous domains considered here. The results we present show that XCSF can converge much faster than tabular techniques while producing more compact solutions. Our results also suggest that when exploration is less effective in some areas of the problem space, XCSF can exploit effective generalizations to extend the evolved knowledge beyond the frequently explored areas. In contrast, in the same situations, the convergence speed of tabular Q-learning worsens.

This paper describes the simulated car racing competition held in association with the IEEE WCCI 2008 conference. The organization of the competition is described, along with the rules, the software used, and the submitted car racing controllers. The results of the competition are presented, followed by a discussion about what can be learned from this competition, both about learning controllers with evolutionary methods and about competition organization. The paper is co-authored by the organizers and participants of the competition.