Image-Based Identification and Localisation of Changes in Intraoperative Brain Tumour Resection

Conventional brain tumour progression tracking methods inherently assume consistent image acquisition conditions between baseline and follow-up scans. In practice, however, MRI data often suffer from harmonisation issues due to variations in acquisition protocols, scanner settings, or patient positioning. To address this, we propose an image-based approach capable of identifying and localising changes between baseline and follow-up MRI scans without requiring independent registration or extensive pre-processing. Our registration-invariant method leverages the internal feature maps of a Siamese network composed of two segmentation models based on dilated convolutions, allowing it to learn long-context spatial features. These features are used to detect and localise changes directly from the input images, which are then enhanced using panchromatic sharpening to emphasise both high-level structural and pixel-level differences within the resulting change map. When tested on previously unseen, unregistered MRI scans, the method outperformed baseline models in identifying and localising resected tumour tissue ( $F_1=0.64$ and $IoU=0.57$ versus baseline $F_1=0.11$ and $IoU=0.07$ ). This performance remained robust under conditions of augmentation, noise, and misalignment ( $F_1=0.55$ and $IoU=0.45$ ), highlighting its resilience to real-world imaging variability. Overall, the proposed approach demonstrates strong robustness to spatial misalignment and offers a promising solution for tracking tumour progression and regression in clinical environments where reliable registration cannot be guaranteed.