Unsupervised Convolutional Autoencoder Framework for Multimodal Medical Image Fusion in Brain Tumour Diagnosis

Medical image fusion improves diagnosis accuracy and reliability by combining images from several modalities. It is gaining prominence for many clinical applications. This paper implements an unsupervised model to fuse gray-scaled Magnetic Resonance Imaging (MRI) with colored Positron Emission Tomography/Single Positron Emission Computed Tomography (PET/SPECT) medical image fusion to locate tumor-affected portions and dead cells clearly. This paper’s main goal is in determining how well an autoencoder’s encoder component can extract features from MRI and PET/SPECT images of brain tumor problems. The autoencoder’s decoder component then uses the features to reconstruct the fused image. The autoencoders are tuned accordingly to get a low Mean Squared Error (MSE) with good structural similarity. It is trained with the dataset of MRI and PET/SPECT images in the whole brain atlas dataset, Harvard University. Our suggested approach has been objectively assessed using four distinct image assessment metrics: Feature Mutual Index (FMI), Structural Similarity Index Measure (SSIM), gradient-based Quality index (Qab/f) and Visual Information Fidelity Factor (VIFF) are compared to four other methods currently in use. In both subjective and objective assessments, our method has outperformed well compared to the existing methods in comparison