Skip to main content
Login | Suomeksi | På svenska | In English

Browsing by Subject "neuroverkko"

Sort by: Order: Results:

  • Pirinen, Riku (2023)
    In recent years, there has been a surge of interest in the application of machine learning (ML) within the medical field. This development has been made possible by the emergence of high-performance modern computers. Deep learning (DL), a subfield of ML, uses advanced computer algorithms to solve complex tasks that would normally require the cognitive capabilities of the human brain. Meningiomas are the most common intracranial tumor excluding metastases. They are usually diagnosed on head magnetic resonance imaging (MRI). Meningiomas can also be visible on computed tomography (CT) although detecting them on non-contrast enhanced head CT is challenging. Notably, CT is more readily available than MRI and head CTs far outnumber head MRIs. This means many meningiomas are not detected on head CT. DL has shown capability in tasks that are difficult for humans, and we hypothesized that a DL algorithm could detect meningiomas from non-contrast-enhanced head CTs. In this thesis, we developed a deep learning algorithm for the detection of clinically relevant meningiomas in non-contrast-enhanced head CT scans and assessed its performance on previously unseen data. All images were from patients who required neurosurgical intervention for the treatment of their meningiomas. They were sourced from the Helsinki University Hospital (HUH) Picture Archiving and Communication System (PACS), a database of over 21 million imaging studies. We found that our algorithm has a good positive predictive value in detecting meningiomas. In other words, the algorithm rarely predicts abnormal findings in normal head CT scans. On the other hand, a positive prediction is likely to be correct. Such an algorithm, in theory, could aid radiologists and clinicians in image interpretation especially during busy on-call hours as it gives few false alerts.
  • Pirinen, Riku (2023)
    In recent years, there has been a surge of interest in the application of machine learning (ML) within the medical field. This development has been made possible by the emergence of high-performance modern computers. Deep learning (DL), a subfield of ML, uses advanced computer algorithms to solve complex tasks that would normally require the cognitive capabilities of the human brain. Meningiomas are the most common intracranial tumor excluding metastases. They are usually diagnosed on head magnetic resonance imaging (MRI). Meningiomas can also be visible on computed tomography (CT) although detecting them on non-contrast enhanced head CT is challenging. Notably, CT is more readily available than MRI and head CTs far outnumber head MRIs. This means many meningiomas are not detected on head CT. DL has shown capability in tasks that are difficult for humans, and we hypothesized that a DL algorithm could detect meningiomas from non-contrast-enhanced head CTs. In this thesis, we developed a deep learning algorithm for the detection of clinically relevant meningiomas in non-contrast-enhanced head CT scans and assessed its performance on previously unseen data. All images were from patients who required neurosurgical intervention for the treatment of their meningiomas. They were sourced from the Helsinki University Hospital (HUH) Picture Archiving and Communication System (PACS), a database of over 21 million imaging studies. We found that our algorithm has a good positive predictive value in detecting meningiomas. In other words, the algorithm rarely predicts abnormal findings in normal head CT scans. On the other hand, a positive prediction is likely to be correct. Such an algorithm, in theory, could aid radiologists and clinicians in image interpretation especially during busy on-call hours as it gives few false alerts.
  • Erälaukko, Hannu (2022)
    Ilmataistelussa on keskeistä pystyä laskemaan, missä kulkee raja sille, milloin kohteen on mahdollista altistua vihollisen ohjusuhalle eli missä sijaistee kohteen ja vihollisen välissä olevan maantieteellisen LAR-alueen (Launch Acceptability Region) raja. Tämä raja voidaan laskea ohjuksen lentämistä simuloivalla simulointiohjelmistokomponentilla. Tämän tutkielman tavoitteena oli löytää ratkaisu, joka mahdollistaisi rajojen sijaintien laskemisen eräässä reaaliaikasovelluksessa siten, että laskeminen perustuisi tällaisen simulointiohjelmistokomponentin tuloksiin, vaikka komponentti itse on liian hidas käytettäväksi kyseisessä reaaliaikasovelluksessa. Päätettiin, että simulointiohjelmistokomponentti korvattaisiin koneoppimismenetelmällä, joka on opetettu jäljittelemään simulointiohjelmistokomponenttia. Ratkaisun löytämiseksi perehdyttiin erilaisiin koneoppimismenetelmiin, joista yksi valittiin alustavaksi ratkaisuksi. Valitun koneoppimismenetelmän, neuroverkon, teoriaan perehdyttiin kirjallisuuskatsauksella, jotta sen kehittämisen tueksi saatiin tietämystä. Neuroverkosta kehitettiin lopullinen ratkaisu Design Science Research -prosessilla. Neuroverkko osoittautui riittävän nopeaksi, että sitä voidaan käyttää halutussa reaaliaikasovelluksessa. Neroverkon kyky jäljitellä simulointiohjelmistokomponentin tuloksia osoittautui myös riittävän tarkaksi.