Anuj Gupta, Anita -, Manish Gupta
The clinically feasible explanations and the high performance in classifications are not only strong screening outcomes of the Magnetic Resonance Imaging (MRI) on brain tumors, but also trustworthy confidence estimates and explanations. The present paper gives a single-run benchmark of three convolutional neural networks (VGG16, ResNet50, EfficientNetB0) and two families of vision transformers (ViT-Base and Swin-Base) in a 100-epoch training scheme in four-class brain MRI classification (glioma, meningioma, pituitary, and no-tumor). It has accuracy, macro-precision/recall/F1, micro-average ROC: AUC, per-class, and confusion matrix reports. In order to measure reliability, it determines calibration by on-assessment of reliability diagrams and Estimated Calibration Error (ECE). To evaluate the interpretability, Grad-CAM images are given to all five models using post-processing (thresholding and smoothing) to identify class-discriminative areas. In addition to discrimination, the assessment presents the differences in confidence, reliability and the quality of explanation among architects as noteworthy, indicating that the accuracy level alone is insufficient to use a system in clinical practice. The outcomes of this work bring forward the need to be reliability-conscious and responsive towards explainability validation before the process is transformed into real neuroimaging practices. The modern backbones (EfficientNetB0 and Swin) are more discriminative and better calibrated models than VGG16, and saliency explainability is also more correlated with pathology, as opposed to classical baselines. It puts the findings in context of the more current state-of-the-art, in terms of summarizing those recently advanced variants of transformer that can be tuned to better calibration output (up to 98.9% accuracy and ECE as low as 0.023) when using multi-scale patch embeddings, feature calibration modules and selective attention.