Jill Selesa Sunday

ABSTRACT

In resource-constrained Sub-Saharan Africa (SSA), the utilization of lower-quality MRI technology raises questions about the applicability of machine learning (ML) methods for clinical tasks. The challenges associated with inadequate image contrast and resolution in SSA-acquired brain MRI scans demand advanced pre-processing methods to improve the image quality before employing ML techniques for tasks like tumor segmentation and classification.

Addressing this, the aim of the BRaTS-Africa 2023 challenge was to provide a platform for research and to come up with a generalizable and robust deep learning-based brain tumor segmentation method tailored for the SSA population.

This study presents a threefold approach. First, the impact of domain shift from SSA training data on model efficacy is highlighted, revealing a discernible influence despite the small dataset size. Secondly, a comparative analysis of 3D and 2D full-resolution models using the nnU-Net version 2 framework indicates the superiority of a 2D model trained for 300 epochs achieving a five-fold cross-validation score of 0.9278. Finally, with respect to the performance gap observed in SSA validation as opposed to GLI validation, two strategies are proposed: fine-tuning SSA cases using the GLI+SSA best pre-trained 2D full resolution model at 300 epochs, and introducing a novel neural style transfer-based data augmentation technique for the SSA cases. The fusion of these strategies significantly improves SSA validation results within computational limitations.

This investigation underscores the potential of notable performance improvements in enhancing brain tumor prediction within SSA’s unique healthcare setting.