Automated Multimodal Fusion Based Hyperparameter Tuned Deep Learning Model for Brain Tumor Diagnosis

Abstract

As medical image processing research has progressed, image fusion has emerged as a realistic solution, automatically extracting relevant data from many images before fusing them into a single, unified image. Medical imaging techniques, such as Computed Tomography (CT), Magnetic Resonance Imaging (MRI), etc., play a crucial role in the diagnosis and classification of brain tumors (BT). A single imaging technique is not sufficient for correct diagnosis of the disease. In case the scans are ambiguous, it can lead doctors to incorrect diagnoses, which can be unsafe to the patient. The solution to this problem is fusing images from different scans containing complementary information to generate accurate images with minimum uncertainty. This research presents a novel method for the automated identification and classification of brain tumors using multi-modal deep learning (AMDL-BTDC). The proposed AMDL-BTDC model initially performs image pre-processing using bilateral filtering (BF) technique. Next, feature vectors are generated using a pair of pre-trained deep learning models called EfficientNet and SqueezeNet. Slime Mold Algorithm is used to acquire the DL models’ optimal hyperparameter settings (SMA). In the end, an autoencoder (AE) model is used for BT classification once features have been fused. The suggested model’s superior performance over other techniques under diverse measures was validated by extensive testing on the benchmark medical imaging dataset.