Detection and Classification of Tumor Type From Brain MRI Images Using Transfer Learning

Abstract

Histopathological analysis of the extracted biological specimen has been one of the most trusted techniques to detect brain tumors in medical diagnostics. However, this analytical approach is invasive, time–intensive, and requires manual intervention; therefore, the probability of manual or human error is high. These practical limitations lay the foundation for identifying a non-invasive and automatic approach to brain tumor detection. Various effective modalities like MRI and CT scans have been discovered. These advancements have aided in gathering preliminary information in case of any suspicions of tumor manifestation. However, a diagnostic conclusion is reached by the subjective evaluation of the medical experts based on the medical images. This again raises the probability of misdiagnosis and, thus, requires an automated diagnostic system that may pitch in a ‘second opinion’ to reduce human error significantly. Deep learning algorithms tend to provide a solution by aiding in the designing of such computer-aided diagnostic systems. Taking this cue, brain tumor detection and classification through EfficientNet-B2 architecture, along with transfer learning, has been presented in the proposed work. Performance analysis of the model has been done by applying transfer learning through ImageNet and Noisy-student and different optimizers on two publicly available datasets. Preliminary results show that an accuracy of 97% is achieved when EfficientNet-B2 is used for tumor classification, which is higher than other models, such as EfficientNetV2B1 (89.17%) and EfficientNetB0 (91%). Also, it is suggestive that noisy student can prove to be an alternative for ImageNet in transfer learning mainly when binary data is being processed.