Automated machine learning-based neural network for brain tumor classification

Abstract

In the contemporary medical landscape, characterized by the widespread and pernicious affliction of brain tumors, the imperative to enhance diagnostic modalities is paramount, aligning with the overarching objective of precisely delineating the affected patient cohort, thereby affording the prospect of administering expeditious therapeutic interventions. Within this scholarly discourse, the principal objective pertains to the discernment of an optimal Convolutional Neural Network (CNN) model, engendered through the mechanizations of automated machine learning, as instantiated within the virtual precincts of the "Edge Impulse" online platform. The corpus of investigation entails the acquisition of requisite data sets from the digital repository denominated "Kaggle," specialized in the domain of scientific knowledge. The amassed data sets, having undergone meticulous preprocessing procedures, were subsequently subjected to partitioning activities within the confines of the "Edge Impulse" framework, wherein a standardized ratio of division, namely a four-to-one proportionality between training and testing subsets, was consistently maintained across discrete data clusters. The training and testing processes were accomplished on Edge Impulse. The image mode, data learning block, learning rate, et cetera, were modified for each neural network models trained on Edge Impulse. Each model performed differently, and the distinct testing accuracy and on device performances were collected for each model for comparison. The experimental results demonstrate that using transfer learning supported by Edge Impulse with learning rate equals to 0.00051 and fit longest axis image resize mode is the optimal option for training a brain classifying model on Edge Impulse through automated machine learning.