BEGIN:VCALENDAR VERSION:2.0 PRODID:-//CERN//INDICO//EN BEGIN:VEVENT SUMMARY:Tomographic reconstruction from proton CT list-mode data using aut omatic differentiation DTSTART;VALUE=DATE-TIME:20260910T081000Z DTEND;VALUE=DATE-TIME:20260910T084000Z DTSTAMP;VALUE=DATE-TIME:20260922T173908Z UID:indico-contribution-421-2069@events.ncbj.gov.pl DESCRIPTION:Speakers: Aurélien Coussat ()\nProton CT is a promising alter native to X-ray CT for proton therapy treatment planning\, allowing direct estimation of the relative stopping power map within the patient body wit hout relying on conversion from Hounsfield units. Conventional list-mode p roton CT scanners measure the energy loss of each individual proton to est imate the integral of the relative stopping power along its path\, the wat er-equivalent path length\, but do not meet the requirements for clinical use due to their low acquisition rates. An alternative proton CT scanner d esign was recently proposed where the time-of-flight of each proton is mea sured between two detectors located before and after the patient along the proton beam. The main advantages of this sandwich time-of-flight design a re its compactness and high acquisition rates. However\, conversion from s andwich time-of-flight data to water-equivalent path length is not possibl e. We present an iterative algorithm that leverages PyTorch's automatic di fferentiation engine to directly optimize the voxels in the image space. T he method is assessed and compared using Monte Carlo simulations.\n\nhttps ://events.ncbj.gov.pl/event/468/contributions/2069/ LOCATION: URL:https://events.ncbj.gov.pl/event/468/contributions/2069/ END:VEVENT BEGIN:VEVENT SUMMARY:Rethinking the Scientific Workflow in the Age of LLMs: Evidence Sy nthesis for Medicine and Beyond DTSTART;VALUE=DATE-TIME:20260910T070000Z DTEND;VALUE=DATE-TIME:20260910T074000Z DTSTAMP;VALUE=DATE-TIME:20260922T173908Z UID:indico-contribution-421-2058@events.ncbj.gov.pl DESCRIPTION:Speakers: Yufang Hou (IT:U)\nhttps://events.ncbj.gov.pl/event/ 468/contributions/2058/ LOCATION: URL:https://events.ncbj.gov.pl/event/468/contributions/2058/ END:VEVENT BEGIN:VEVENT SUMMARY:THE APPLICATION OF REINFORCED MACHINE LEARNING FOR CACHE MANAGEMEN T IN HIGH-LOAD INFORMATION SYSTEMS DTSTART;VALUE=DATE-TIME:20260910T074000Z DTEND;VALUE=DATE-TIME:20260910T081000Z DTSTAMP;VALUE=DATE-TIME:20260922T173908Z UID:indico-contribution-421-2036@events.ncbj.gov.pl DESCRIPTION:Speakers: Bohdan Volokh (Kyiv National University of Construct ion and Architecture)\nThe performance of modern\, high-load information s ystems significantly depends on the efficiency of cache management. Meanwh ile\, the traditional Least Recently Used and Least Frequently Used polici es are based on fixed heuristics\, which limits their ability to adapt to changes in query structure and intensity.\n\nTo overcome these constraints \, a decision-making approach for adding new objects to the cache is propo sed\, which is based on the tabular Q-Learning method with a deferred asyn chronous update mechanism. It takes into account the key characteristics o f objects\, the context of their use and long-term consequences. The propo sed approach involves training the agent through the accumulation of exper ience and the consideration of deferred rewards\, which ensures an adaptiv e caching policy in real time and improves its efficiency\, in particular by increasing the cache hit ratio.\n\nThe practical value of the results l ies in the ability to improve the efficiency of Clinical Decision Support System users when working with current data. These systems process large\, rapidly updated sets of medical data\, and the developed approach will re duce data access latency and improve the information support provided to d octors when forming clinical recommendations. The proposed approach can al so be applied to other electronic health systems for making decisions rega rding the caching of frequently used data.\n\nhttps://events.ncbj.gov.pl/e vent/468/contributions/2036/ LOCATION: URL:https://events.ncbj.gov.pl/event/468/contributions/2036/ END:VEVENT BEGIN:VEVENT SUMMARY:Machine Learning for Event-Level Background Rejection in LAFOV PET DTSTART;VALUE=DATE-TIME:20260909T094000Z DTEND;VALUE=DATE-TIME:20260909T101000Z DTSTAMP;VALUE=DATE-TIME:20260922T173908Z UID:indico-contribution-421-2051@events.ncbj.gov.pl DESCRIPTION:Speakers: Michał Obara (NCBJ)\nLong axial field-of-view PET s canners provide a significant gain in sensitivity\, but also bring a subst antially higher background of scattered and accidental coincidences. We in vestigate machine-learning classification of individual coincidence events using GATE Monte Carlo simulations of the Siemens Quadra scanner with NEM A IEC and anthropomorphic XCAT phantoms\, where every event has a ground-t ruth label.\n\nWe evaluate XGBoost\, AdaBoost\, and neural network classif iers with two feature sets\, and show that phantom-wise metrics give a mis leading picture of classifier performance. In a cross-phantom test\, a sma ll decrease in accuracy corresponds to regionally concentrated degradation that only per-voxel quality maps reveal\, while geometry-dependent featur es overfit the phantom geometry.\n\nIn our initial approach\, we kept even ts predicted as true and rejected all others. We present two extensions i n which the per-event probabilities influence the reconstructed image and its uncertainty. First\, using soft probability weighting\, in which each event contributes to reconstruction according to its estimated true-coinci dence probability\, we reduced the image-wise reconstruction error compare d to hard filtering. Second\, we apply split conformal prediction\, which replaces a single predicted class with a set of classes carrying a finite- sample coverage guarantee. Aggregated per voxel\, these sets produce ambig uity maps that identify unreliable image regions without ground-truth labe ls.\n\nhttps://events.ncbj.gov.pl/event/468/contributions/2051/ LOCATION: URL:https://events.ncbj.gov.pl/event/468/contributions/2051/ END:VEVENT BEGIN:VEVENT SUMMARY:Overview and current challenges of practical CT reconstruction DTSTART;VALUE=DATE-TIME:20260909T083000Z DTEND;VALUE=DATE-TIME:20260909T091000Z DTSTAMP;VALUE=DATE-TIME:20260922T173908Z UID:indico-contribution-421-2047@events.ncbj.gov.pl DESCRIPTION:Speakers: Ander Biguri (University of Cambridge)\nData-driven reconstruction for computed tomography has boomed in the last decade. Yet unlike in other AI fields\, its adoption in practical CT has been slow. I n this talk I'll give a brief overview of the types of reconstruction meth ods and highlight the current challenges and some progress that we have do ne recently\n\nhttps://events.ncbj.gov.pl/event/468/contributions/2047/ LOCATION: URL:https://events.ncbj.gov.pl/event/468/contributions/2047/ END:VEVENT BEGIN:VEVENT SUMMARY:A Fast and Generic Energy‑Shifting Transformer for Hybrid Monte Carlo Radiotherapy Dose Calculation DTSTART;VALUE=DATE-TIME:20260909T091000Z DTEND;VALUE=DATE-TIME:20260909T094000Z DTSTAMP;VALUE=DATE-TIME:20260922T173908Z UID:indico-contribution-421-2021@events.ncbj.gov.pl DESCRIPTION:Speakers: Chi-Hieu Pham (LaTIM)\, Julien Bert (LaTIM\, UMR1101 INSERM)\nWe introduce a novel learning framework for accelerated Monte Ca rlo (MC) dose calculation termed Energy Shifting. This approach leverages deep learning to synthesize complex polyenergetic dose distributions direc tly from simple monoenergetic inputs under identical beam configurations. Unlike conventional approaches such as denoising or GenAI techniques\, our method achieves superior cross-domain generalization on unseen datasets b y integrating high-fidelity anatomical textures and source-specific beam s imilarity directly into the model’s input space. Furthermore\, we propos e a novel 3D architecture termed TransUNetSE3D\, featuring Transformer blo cks for global context and Residual Squeeze-and-Excitation modules for ada ptive channel-wise feature recalibration. Hierarchical representations of these blocks are fused into the network’s latent space alongside the pri mary dose-map parameters\, allowing physics-aware reconstruction. This hyb rid design outperforms existing U-Net and Transformer-based benchmarks in both spatial precision and structural preservation\, while maintaining the execution speed necessary for real-time use. Our proposed pipeline achiev es a Gamma Passing Rate exceeding 98% (3%/3mm) compared to the MC referenc e\, evaluated within the framework of a treatment planning system using 6M V TrueBeam Linear Accelerator for prostate radiotherapy. These results off er a robust solution for fast volumetric dosimetry in adaptive radiotherap y.\n\nhttps://events.ncbj.gov.pl/event/468/contributions/2021/ LOCATION: URL:https://events.ncbj.gov.pl/event/468/contributions/2021/ END:VEVENT BEGIN:VEVENT SUMMARY:Graph Neural Networks and Multi-Objective Optimization for Coded-M ask Prompt-Gamma Imaging within the SiFi-CC Collaboration DTSTART;VALUE=DATE-TIME:20260909T101000Z DTEND;VALUE=DATE-TIME:20260909T104000Z DTSTAMP;VALUE=DATE-TIME:20260922T173908Z UID:indico-contribution-421-2032@events.ncbj.gov.pl DESCRIPTION:Speakers: Philippe Clement (Doctoral Schoool of Exact and Natu ral Sciences\, Jagiellonian Universtity\, Kraków)\n**Background and Aims: ** This study introduces graph neural networks (GNNs) to improve prompt-ga mma imaging for a 1D coded-mask gamma camera developed by the SiFi-CC coll aboration for proton therapy verification. \n\n**Methods:** The coded-mask detector features LYSO:Ce\,Ca scintillating fibers with dual-ended readou ts. Coincident detector signals are aggregated into event representations and subsequently into cluster representations to capture spatial\, tempora l\, and energy information. A graph-based learning model infers interactio n positions and energies from these representations\, feeding the output i nto a Maximum-Likelihood Expectation-Maximization (MLEM) algorithm for dep th profile reconstruction. Finally\, image reconstruction parameters are o ptimized using an NSGA-III multi-objective genetic algorithm within the Op tuna framework.\n\n**Results:** For a dataset of 10$^{10}$ protons\, the o ptimized pipeline demonstrated sub-millimeter beam range accuracy\, achiev ing a root-mean-square error (RMSE) below 0.6 mm. \n\n**Conclusions:** Com bining graph-based event reconstruction with targeted image optimization s ignificantly improves the resolution of coded-mask prompt-gamma imaging de tectors.\n\nhttps://events.ncbj.gov.pl/event/468/contributions/2032/ LOCATION: URL:https://events.ncbj.gov.pl/event/468/contributions/2032/ END:VEVENT BEGIN:VEVENT SUMMARY:Diffusion Priors for Physics-Informed CT and PET Reconstruction DTSTART;VALUE=DATE-TIME:20260909T070000Z DTEND;VALUE=DATE-TIME:20260909T074000Z DTSTAMP;VALUE=DATE-TIME:20260922T173908Z UID:indico-contribution-421-2042@events.ncbj.gov.pl DESCRIPTION:Speakers: Alexandre Bousse (LaTIM U1101)\nGenerative models pr ovide powerful learned priors for solving challenging inverse problems in medical imaging. In this talk\, I will present diffusion-based\, physics-i nformed approaches for CT and PET reconstruction\, illustrated through thr ee applications: motion-compensated head cone-beam CT\, material decomposi tion in photon-counting CT\, and joint activity-attenuation reconstruction in CT-less PET. These examples demonstrate how learned image priors can b e combined with measurement physics to achieve high-quality\, data-consist ent reconstruction without paired training data.\n\nhttps://events.ncbj.go v.pl/event/468/contributions/2042/ LOCATION: URL:https://events.ncbj.gov.pl/event/468/contributions/2042/ END:VEVENT BEGIN:VEVENT SUMMARY:Run in seconds on your laptop: A lightweight model for CT segmenta tion of maxillary sinuses in children with chronic rhinosinusitis DTSTART;VALUE=DATE-TIME:20260909T074000Z DTEND;VALUE=DATE-TIME:20260909T081000Z DTSTAMP;VALUE=DATE-TIME:20260922T173908Z UID:indico-contribution-421-2023@events.ncbj.gov.pl DESCRIPTION:Speakers: Przemysław Olbratowski (Biological and Chemical Res earch Centre\, University of Warsaw)\nWe present a compact 3D neural netwo rk for segmentation of maxillary sinuses from CT scans in children with ch ronic rhinosinusitis. It comprises 12 convolutional layers totaling 50k pa rameters\, minimal compared to most 3D medical-imaging architectures\, and completes the prediction in a few seconds on an average CPU\, making the tool accessible for most clinicians. The sinus volume and its fraction occ upied by inflammatory changes are estimated with an RMSE of 0.5 cm$^3$ and 1.5 %pts\, respectively\, which allows for precise longitudinal monitorin g. The network includes no dropout\, no pooling\, and no padding\, the lat ter preventing layer-wise injection of meaningless zeros. We utilize a cus tom normalization that does not average the data but employs the same runn ing statistics in both prediction and training\, making the model insensit ive to overall contrast and local artifacts. The network is fully convolut ional and translation-invariant. It has an inner receptive field of 18x18x 18 voxels to detect sinus walls and an outer receptive field of 88x88x88 v oxels to provide a broader context. The internal data flow is designed to minimize the number of mappings that the network must learn. It operates i n a reduced resolution of 1x1x1 mm$^3$\, which still allows for high preci sion thanks to the use of fuzzy labels accounting for the partial-volume e ffect. The model was trained and tested on a dataset of 92 scans collected and manually annotated specifically for this study.\n\nhttps://events.ncb j.gov.pl/event/468/contributions/2023/ LOCATION: URL:https://events.ncbj.gov.pl/event/468/contributions/2023/ END:VEVENT END:VCALENDAR