Correlation-based feature enhancement for Non-Enhancing brain tumor segmentation – Sharjeel Masood (Extra Seminar, Online)

Brain tumor datasets typically contain four MRI modalities—T1-weighted (T1), T1-weighted with contrast enhancement (T1c), T2-weighted (T2), and Fluid-Attenuated Inversion Recovery (FLAIR)—to provide a comprehensive view of the tumor and surrounding tissues. Each modality highlights different properties of the brain, and are used together to segment 3 types of tumors, Enhancing tumors, Non-Enhancing tumors and Edema. The main problem that most segmentation algorithms face is that both non-enhancing tumors and edema are characterized by increased fluid content, which makes them appear as hyperintensities (bright areas) on T2-weighted and FLAIR sequences. Slight variations in the quality of the scan or noise can make edema and non-enhancing tumors indistinguishable (which is common), this then needs to be solved by human intervention using contextual clues like location, shape or patient history.

I will be presenting a feature enhancement module with the ability of enhancing the information in each individual slice by leveraging the contextual information from adjacent slices. It uses a correlation module that acts as a trainable feature selector and decides which features in the input image need to be enhanced. This is a computationally inexpensive enhancement module that can be attached with a segmentation model to improve it's performance on non-enhancing tumors and edema. The presentation would also discuss results of how the enhancement module performs as the visibility of these tumors decreases.