BEGIN:VCALENDAR VERSION:2.0 PRODID:-//CERN//INDICO//EN BEGIN:VEVENT SUMMARY:Procedural generation of tomographic datasets for machine learning DTSTART;VALUE=DATE-TIME:20260909T123000Z DTEND;VALUE=DATE-TIME:20260909T130000Z DTSTAMP;VALUE=DATE-TIME:20260922T182732Z UID:indico-contribution-2038@events.ncbj.gov.pl DESCRIPTION:Speakers: MikoĊ‚aj Mrozowski (NCBJ)\nIndustrial computed tomog raphy allows for non-destructive inspection of complex components\,\nmakin g it an interesting domain of applications for machine learning methods.\n In this context\, a large bottleneck for training robust models is data ac quisition.\nAlready time-intensive task of obtaining data using real-world CT scanners \nis further exacerbated by the requirement of creating accur ate labels.\n\nWe solve this problem by implementing a pipeline for the ge neration of massive \nsynthetic datasets\, and using only some examples of real-world data.\n\nOur data pipeline uses parametrized CAD geometry to g enerate thousands of \nrandomized objects from which we derive both ground -truth labels and \nattenuation data further fed into tomographic simulati ons.\n\nThe simulation code is GPU-accelerated allowing us to substantiall y \nreduce the time needed to generate the datasets.\n\nTo reproduce effec ts visible in real data\, we implement a polychromatic beam\,\nand simulat e artefacts such as non-uniformity of the beam and detector matrix respons e.\nWe also apply custom beam hardening corrections for our reconstruction s.\n\nWe use the synthetic data in semi-supervised training of a W-net mod el for the specific\ntask of segmentation of ionization chambers. We show\ , that semi-supervised approach \nallows for effective training with only partial labeling (as low as 25%-50% of the data).\nWe then validate perfor mance of segmentation models trained using simulations on the real-world d ata.\n\nhttps://events.ncbj.gov.pl/event/468/contributions/2038/ LOCATION: URL:https://events.ncbj.gov.pl/event/468/contributions/2038/ END:VEVENT END:VCALENDAR