Speaker
Description
Paweł Konieczka, on behalf of the J-PET collaboration
Convolutional Neural Networks are excellent at analyzing images by
learning abstract representations. CNN has been an overwhelming strategy
in computer vision tasks and has achieved expert-level performances in
various fields. There has been a surge of interest in the potential of
CNN among radiology researchers and several studies have already been
published in areas such as classification [1] and image reconstruction
[2].
First general methodology to transform a non-image data into an image
for CNN architectures has been presented in [3]. Nevertheless, this
method cannot be applied to large data sets, where number of features is
very small, because of computational complexity of PCA. The introduction
of scheme of non-image data transformation into 2-dimensional matrices
will be proposed [4].
The goal of this poster is to present results of multi-photon
coincidences classification in J-PET scanner using CNNs. Bayesian
optimization of two convolutional network architectures (DeepInsight
[3], YOLOv1 [5]) will be presented.
References:
[1] Yasaka, Koichiro, et al. Deep learning with convolutional neural network
for differentiation of liver masses at dynamic contrast-enhanced CT: a
preliminary study. Radiology, 2018, 286.3: 887-896.
[2] Liu, Fang, et al. Deep learning MR imaging--based attenuation correction
for PET/MR imaging. Radiology, 2018, 286.2: 676-684.
[3] Sharma, Alok, et al. DeepInsight: A methodology to transform a non-image
data to an image for convolution neural network architecture. Scientific
reports, 2019, 9.1: 1-7.
[4] Raczyński, Lech, Introduction of non-image PET data transformation to
image-form approach for classification using Convolutional Neural
Networks. 1st Symposium on Theranostics, 2021.
[5] Redmon, Joseph, et al. You only look once: Unified, real-time object
detection. In: Proceedings of the IEEE conference on computer vision and
pattern recognition. 2016. p. 779-788.
E-mail: pawel.konieczka@ncbj.gov.pl
