Hybrid deep learning network enabled brain tumor detection with shape-aware loss-based structure correcting adversarial network using MRI

The brain plays the most critical role in the human body. Brain tumors are a major health problem around the world. These tumors are caused when abnormal cells grow inside the brain. Identifying brain tumors at an early stage greatly enhances treatment success and survival rates.The ability to detect tumors quickly and accurately is key to better treatment results. Magnetic Resonance Imaging (MRI) scans can automatically find these tumors, but this is a difficult task because the visual appearance of tumor tissue is very similar to healthy brain tissue. This similarity often leads to mistakes in diagnosis and false positives. Existing methods face limitations in feature discrimination, segmentation accuracy, and robustness against image noise. In this research, MRI-based early brain tumor detection is performed using the proposed Deep Higher Order Kronecker Forward Harmonic Network (DHKFHNet). Initially, MRI input images are subjected to denoising with the help of a Midpoint filter that maintains important image details while effectively minimizing noise. Followed by denoising, the process of skull removal by the Brain Surface Extractor (BSE) algorithm is done, which isolates the brain tissue from non-brain structures. By removing irrelevant data and noise, the model gives more importance to the brain structure, thereby providing better outcomes. Next, tumor area segmentation is done by the Shape-aware loss-based Structure Correcting Adversarial Network (SCAN). The segmentation process isolates the tumor from healthy tissue, thereby reducing the chance of misclassification. Then, the necessary features essential for tumor detection are extracted. By focusing on relevant features, the devised models can learn more effectively, thereby providing better detection and diagnosis results. In the final stage of the proposed framework, brain tumor detection is carried out using the DHKFHNet model, developed by combining the Deep Higher Order Attention Neural Network (DHA-Net), the Deep Kronecker Network (DKN), and harmonic principles. The datasets, namely Br35H: Brain Tumor Detection 2020, Brain Tumor Classification (MRI), and BraTS 2021 Task 1 are used for evaluating the performance of the proposed model. Moreover, the proposed DHKFHNet model is analyzed based on the accuracy, sensitivity, specificity, and F1-score metrics, which achieved superior results of 92.99%, 92.97%, 91.89%, and 92.55%, respectively. The proposed DHKFHNet model gained performance improvement of 7.66%, 6.43%, 4.88%, 3.73%, 2.65%, 2.55%, and 2.10% higher than the State-of-the-Art (SOTA) methods when considering the accuracy metrics for dataset 2. These findings of the DHKFHNet demonstrated that the model highly supports radiologists and clinicians, which helps to detect brain tumor with improved accuracy and reliability, contributing to improved clinical decisions.