This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
,
Faculty of physics, University of Belgrade , Belgrade , Serbia
Institute of Oncology and Radiology of Serbia, Institute of Oncology and Radiology of Serbia , Belgrade , Serbia
Faculty of physics, University of Belgrade , Belgrade , Serbia
Clinic for Radiation Oncology, Institute of Oncology and Radiology of Serbia , Belgrade , Serbia
Faculty of Medicine, University of Belgrade , Belgrade , Serbia
This study investigated the application of neural networks for detecting rectal tumor regions in patients using computed tomography (CT) images. The study included 138 patients, while two CT scans with a healthy rectum were included to balance the dataset, resulting in a total of 3566 images. For each image, pixel values along the contour and within the rectal region were extracted, and tissue mass density values were obtained using calibration curves. To account for variations in rectal volume, statistical distribution functions of tissue density were used. The most relevant features were selected using Mutual Information (MI) and Principal Component Analysis (PCA). These data were then used to train different neural network architectures, including layer configurations such as n–2n–n–1 and n–2n–4n–2n–n–1, where n represents the number of input features, which ranged from three to seven. We trained the networks with batch sizes of 1, 11, 107, and 1069, and evaluated their performance based on sensitivity and specificity.We observed the highest accuracy (~79%) with a 4–8–4–1 network, batch size 1069, and 4 PCA-selected features. The highest sensitivity (~85%) was achieved with 6–12–24–12–6–1 and 4–16–8–4–1 networks (batch size 1), using 6 PCA-selected features in the first and 4 MI-selected features in the second. We recorded the highest specificity (~80%) for the 4–8–4–1 network with batch size 1069 and 4 PCA-selected features.
MM is supported by the Horizon Europe STEPUPIORS Project (HORIZON-WIDERA-2021-ACCESS-03, European Commission, Agreement No. 101079217) and the Ministry of Science, Technological Development and Innovation of the Republic of Serbia (Agreement No. 451-03-136/2025-03/200043). ED utilized computational resources provided by the National Data Centre of the Republic of Serbia.
The statements, opinions and data contained in the journal are solely those of the individual authors and contributors and not of the publisher and the editor(s). We stay neutral with regard to jurisdictional claims in published maps and institutional affiliations.
