Abstract
Bladder tumours (BTs) pose significant clinical challenges due to their high recurrence rates and risk of progression to invasive malignancies, which emphasises the need for early and accurate detection. Magnetic resonance imaging (MRI), with its superior soft tissue contrast, is a potential modality for BT detection. To analyse the MRI scans, artificial intelligence (AI) models are increasingly being leveraged. However, these models are often limited by a scarcity of annotated datasets, challenges in pixel-level tumour prediction, and insufficient transparency in predictions. This study introduces the Explainable and Likelihood-Aware AI (ELAAI) framework, designed to address these limitations. Trained solely on annotated normal bladder MRI scans, ELAAI integrates three novel modules: MFA-Net, a robust multi-scale feature aggregation network for bladder segmentation; a refinement step employing adaptive tolerance technique to enhance segmentation of irregularities; and a single-step likelihood prediction network (SLIP-Net), which is a vision transformer with a novel multi-scale deterministic uncertainty (MSDU) head for tumour likelihood prediction. Rigorous evaluation against state-of-the-art (SOTA) models highlights ELAAI's superior performance, enhancing transparency, and reliability in clinical settings by fostering trust in AI-assisted decision-making.