A HYBRID CNN-TRANSFORMER NETWORK FOR THE SEGMENTATION OF HIPPOCAMPUS FROM BRAIN MRI

Abstract

            The physical features of Hippocampus (hc) in human brain indicate several neurological disorders such as Alzheimer. Magnetic Resonance Images (MRI) play a prominent role in radiotherapy planning. Hippocampus segmentation methods for radiotherapy planning make use of brain atlas, convolution neural network etc. We proposed a strategic CNN model with transformer layer which includes several human-engineered feature filters to segment hippocampus. We investigate the performance of the proposed features set in segmenting hippocampus from MRI without make much loss in detection. The feature set includes contrast of the MRI, wavelet implemented frequency components, energy level of grey values and regular statistical parameters. The features are provided as input to a deep learning model and ensured the opted activation function. The qualities are measured using dice, precision, recall, sensitivity, specificity and accuracy. The results ensured the proposed features pioneering the segmentation process.