Development of a Hybrid Deep Learning model with self-adaptive noise resilience for accurate brain tumor detection in MRI scans

Abstract

Medical imaging professionals need brain tumor classification that achieves precise diagnosis and treatment planning while MRI images degrade their classification effectiveness because of contamination by noise. The research presents an automatic deep learning system which modifies its training protocols through real-time noise measurement to boost classification outcomes. MobileNet works together with the SelfAdaptiveConv2D layer that uses adaptive features extraction for better noise distortion resistance. The image quality enhancement process depends on four preprocessing methods: grayscale conversion and noise reduction and CLAHE-based contrast enhancement and normalization techniques. The model receives training and evaluation on the Kaggle Brain Tumor dataset that includes 3,264 MRI images sorted into four earthworm classes: meningioma and glioma in addition to pituitary tumor and no tumor. The proposed model performs considerably better than DenseNet, CNN, and InceptionResNetV2 architecture based on multiple tests which revealed 95% accuracy with precision metrics, recall measurement, F1-score calculation, sensitivity detection and specificity evaluation. The effectiveness of minimizing misclassifications becomes clear through confusion matrix analysis. A wide range of studies demonstrates that this proposed detection method outperforms brain tumor recognition thereby establishing itself as a promising solution for medical diagnostics and automation in the detection of tumors in MRI scans with noise.