Identifying Dataset Faults to Enhance Glioma Classification and Multiclass Brain Tumor Analysis with Deep CNN Architectures

Brain tumor analysis using machine learning models relies heavily on dataset quality. This study analyses the performance of InceptionV3 and VGG16 models on a widely used brain tumor MRI dataset for classification. Initial experiments yielded suboptimal accuracies of 65 percent and 74 percent, respectively, with significant misclassifications in glioma cases. Although previous studies have noted challenges with the glioma class of this corresponding dataset, a detailed analysis was lacking. Further, the models were retrained excluding the problematic glioma class, achieving improved accuracies of 90.63 percent and 96.49 percent for VGG16 and InceptionV3, with InceptionV3 outperforming VGG16. With the help of a neurologist ground truth verification have been performed and the extensive labeling inconsistencies within the glioma class confirmed. Additionally, glioma images from an independent dataset verified by the neurologist were integrated into the experiment under equal preprocessing techniques such as, gaussian blur, CLAHE, normalization along with data augmentation and significant improvement of the experimented models was observed classifying the initial four classes. This study underscores the life-threatening risks of deploying models trained on unverified datasets and highlights the critical importance of rigorous dataset validation in medical imaging. By systematically tackling the problems of data quality, this work not only solves a big challenge in brain tumor classification but also sets a framework for increasing the reliability of machine learning applications in healthcare.