BrainMorphNet (BMN): Multistage Hierarchical Attention Framework for Robust Brain Tumor Detection in MRI
Аннотация
Abstract The development of effective treatments and the improvement of patient survival rates depend heavily on the timely and precise diagnosis of brain tumors. Nevertheless, the current deep-learning methods have trouble generalizing, are prone to overfitting with small datasets, and are imprecise in identifying the subtle boundaries of tumors, all of which could limit their application in clinical settings. To tackle these concerns, this paper has introduced an innovative BrainMorphNet (BMN) structure which can automatically detect and categorize brain tumors on MRI images completely, hence overcoming these problems. The first step of the approach is the application of data-specific preprocessing to clear the images of artifacts and improve their quality. Afterwards, the multilevel feature extraction generates the global and local structures of the tumor regions simultaneously. Lastly, the framework applies adaptive attention mechanisms to spatial and contextual features and thus improve them. There is a very strong multiclass separation between gliomas, meningiomas, pituitary tumors, and nontumor cases, and the whole setup is such that cancers can be diagnosed at a particular place. Our rigorous testing on the Brain Tumor MRI and Figshare datasets gave a model accuracy of 0.993, precision of 0.985, and F1 score of 0.988. The system is a very effective and time-saving tool for brain tumor analysis automation, and it is a significant step towards more intelligent medical imaging and diagnostic systems. These impressive numbers show that BMN has good generalization and clinical reliability.