SU-FTD2: Transformer-Driven Brain Tumor Imaging Framework Using Explainable AI for Consumer Applications

In the clinic, brain tumor detection and semantic segmentation are critical tasks in medical image analysis, particularly in magnetic resonance imaging (MRI), which assists clinicians in diagnosing and monitoring tumor progression. However, convolutional models struggle with long-range information due to limited kernel sizes, which affects the precision of tumor segmentation, especially for tumors of varying sizes. On the other hand, transformer architecture excels at capturing long-range dependencies across domains such as computer vision. This inspired the adoption of vision transformer-based architectures for improved medical image analysis. This study presents SUFTD2, a novel segmentation framework that combines (Swin UNET Transformers) Swin UNETR with a finetuned Detectron2 model to detect and segment brain tumors precisely. Swin UNETR treats brain tumor segmentation as a sequence-to-sequence tumor prediction task, projecting multi-modal image data into a 1D sequence for processing by a hierarchical Swin UNETR encoder. This encoder extracts features at multiple resolutions using shifted windows for self-attention computation. A finetuned model from the Detectron2 model zoo is also used to detect tumor locations and regions. Our suggested framework supports continuous, decentralized medical monitoring by bridging AI-driven segmentation with connected healthcare infrastructure. It incorporates explainable AI (XAI) methods, including LIME and SHAP, to highlight tumor regions that influence predictions, thus improving clinical trust and interpretability. Using XAI and consumer-deployable technologies, the framework advances the accessibility and transparency of brain tumor diagnostics enabled by ubiquitous consumer electronics, aligning with next-generation connected healthcare goals. The proposed XAI-integrated Swin UNETR framework, which leverages consumer electronics for deployment, makes more advanced brain tumor diagnostics available and transparent. The proposed scheme has been tested on the BraTS2020 dataset. We use several performance evaluation matrices to evaluate the experimental outcomes of the proposed scheme. The proposed approach demonstrates outperformance compared to state-of-the-art models.