Classification of the brain metastases based on a new 3D deep learning architecture

• Published in: Soft computing (Berlin, Germany) Vol. 27; no. 22; pp. 17243 - 17256
• Main Authors: Cuşkun, Yasin, Kaplan, Kaplan, Alparslan, Burcu, Ertunç, H. Metin
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.11.2023; Springer Nature B.V
• Subjects: Accuracy; Algorithms; Application of Soft Computing; Artificial Intelligence; Artificial neural networks; Brain; Brain cancer; Breast cancer; Classification; Classification schemes; Communications systems; Computational Intelligence; Control; Data augmentation; Deep learning; Effectiveness; Engineering; Image classification; Image segmentation; Lung cancer; Machine learning; Magnetic resonance imaging; Mathematical Logic and Foundations; Mechatronics; Medical imaging; Melanoma; Metastasis; Recovery (Medical); Robotics; Skin cancer; Success; Support vector machines; Tumors; Wavelet transforms; Deep learning; CNN; Brain metastases; Medical imaging; Classification
• ISSN: 1432-7643; 1433-7479
• DOI: 10.1007/s00500-023-08051-w

Malignant tumors can cause to brain tumors by the spreading to the brain. When the primary site of origin of brain metastases are differentiated by deep learning algorithms, doctors can begin treatment in time without performing biopsy. In this study, we aimed to classify the primary site of brain metastases from the MR images of patients with brain tumors. The MR images were read by radiologists from Kocaeli University Department of Radiology by using their Picture Archiving and Communication System (PACS). We used the three-dimensional convolutional neural networks technique, which is one of the artificial intelligence techniques that are frequently used to determine image classification and identify segmentation problems for the classification using a data set consisting of labeled MR images. Contrary to the models used in the studies performed on MR images in the literature, four different sequences were submitted to the classification model after four different convolutional processes were conducted, although not simultaneously. The classification scheme was designed as a binary classification and a three group classifications. The results showed that 81.48% accuracy rate was obtained in the binary classification and 74.07% accuracy was obtained in the three-group classification. To improve the accuracy rate of the three group classifications, data augmentation was applied and the accuracy rate improved to 75.92%.