Tumor detection in MRI Images Based on Transfer Deep Learning Architectures Driven by Neutrosophic Features

Abstract

MRI images are often noisy, inhomogeneous, and suffer from partial volume effects, which add a great deal of uncertainty to the identification of brain tumors by computer vision techniques. In the present paper, a new fourth domain of Neutrosophic Sets (NS) named Homogeneity (H) has been used for the quantification of the local tissue homogeneity. The proposed framework classifies the images of 600 brain MRI scans into normal or tumor types by automatically applying a hybrid integration of the images of the four domains of the Neutrosophic Sets with the images of the three state-of-the-art deep transfer learning architectures EfficientNet-B0, ResNet50, and DenseNet121, followed by a Support Vector Machine (SVM) classifier. The preprocessing steps are implemented through a robust process that includes median filtering, AHE, min-max normalization, and bicubic interpolation. Among all these networks, EfficientNetB0 has shown the highest classification performance. It has produced an accuracy of 99.8%, a sensitivity of 99.6%, a specificity of 100.0%, a precision of 100.0%, and an F1-score of 99.8% for Falsity (F) domain data using a 70/30 data split. It has shown a performance improvement of 1.7% over the baseline NS+VGG16 and has used 26 times fewer parameters than the baseline (5.3M vs. 138M). In this research, a new Homogeneity (H) domain has been introduced. It has shown a similar performance of 99.8% using the EfficientNetB0 network. ResNet50 has shown a second-highest performance of 99.3% using F-domain data and a 70/30 data split. However, DenseNet121 has shown poor performance compared to the baseline and has produced a poor performance of 95.7%. A new finding is that all these networks are using Falsity (F) domain data more than the Indeterminacy (I) domain data. It is a major change from the baseline VGG16 network that showed a high performance using the Indeterminacy (I) domain.