CT-γ-Net: A Hybrid Model based on Convolutional Encoder-Decoder and Transformer Encoder for Brain Tumor Localization

Abstract

Brain Tumor is a life-threatening disease, and its early diagnosis can save human life. Computer-Aided Brain Tumor Segmentation or Localization in Magnetic Resonance Imaging (MRI) images have emerged as pivotal approaches for expediting the disease diagnosis process. In the past few decades, various researchers combined the strengths of Convolutional Networks and Transformer to perform Brain Tumor Segmentation. However, these models require a large number of trainable weights parameters, and there is still scope for performance improvement in them. To bridge these research gaps, this paper proposes a novel hybrid model named “CT-γ-Net” for effective and efficient Brain Tumor Localization. The proposed CT-γ-Net model follows an Encoder-Decoder structure in which the Convolutional Encoder (CE) and Transformer Encoder (TE) are used for encoding, whereas the Convolutional Decoder (CD) is utilized for decoding the combined output of CE and TE to generate the segmentation masks. In CE and CD components of CT-γ-Net model conventional convolutional layers are replaced by Depth-Wise Separable convolutional layers, as these layers significantly reduce trainable weights parameters. The proposed model achieves 95.5% MeanIoU, 94.82% Dice Score, and 99.24% Pixel Accuracy on publicly available dataset named The Cancer Imaging Archive (TCIA). These experimental results demonstrate that the CT-γ-Net model outperformed other state-of-the-art research works, despite using roughly 28% fewer trainable weights parameters. Hence, the proposed model’s lightweight nature and its high performance, makes it a suitable candidate for deployment on mobile devices, facilitating the precise localization of brain tumor regions in MRI images.