Questions worth following.

Abstract

Medical image analysis using deep learning has made tremendous progress, but the robustness of such systems in realclinical settings is still hampered by model instability and adversarial sensitivity. This paper presents an efficient ensemble-motivated framework for MRI-based brain tumor prediction, incorporating three state-of-the-art networks (including InceptionV3, MobileNetV2 and ResNet50) to enhance diagnosis correctness and prognosis stability. In order to compensate the lack of dataset, we used heavy augmentations such as adding Gaussian noiseand rotation to between fivefold of the amount in training data and obtained the improved generalization. Every model was fine-tuned onaugmented dataset, and their probabilistic predictions were aggregated via soft-voting ensemble learning to achieve more consistent and accurate results than any single model. Apart from general classification, we focus on adversarialthreats to medical imaging systems in this work. We used four popular adversarial attack algorithms—FGSM, BIM, PGD and CW to analyze how easily our model could be fooled by the imperceptible perturbations that are able to corruptclinical judgment. To defend against these threats, adversarial training and ensemble architectural diversity were employed as defenses. Experimental results show adding the proposed technique increases this accuracy to 92% and our particularly higher recall in tumours lead to a lower risk of false-negativediagnoses. In addition, the ensemble is more robustunder adversarial perturbations than each individual model. In general, this workemphasizes the importance of combining high-performance classification and adversarial robustness for safe, reliable, as well as trustworthy AI-based brain tumor detection in clinical care.