Abstract:
"Brain tumour segmentation from MRI data is crucial for accurate diagnosis and treatment planning,
yet it remains challenging due to the heterogeneity and complex spatial structures of tumours. This
paper introduces a novel segmentation approach that combines classical neural networks with quantum
computing elements, aiming to improve accuracy and robustness in delineating tumour boundaries. By
integrating the computational strengths of both paradigms, our method represents a significant
advancement in brain tumour segmentation techniques, offering potential for more precise diagnosis
and personalized treatment in clinical settings.
Utilizing the TransBTS 2020 challenge dataset, known for its extensive collection of annotated MRI
scans, we benchmark our hybrid classical-quantum model against traditional CNN architectures. Our
findings highlight the limitations of conventional CNNs in capturing the intricate spatial relationships
in brain tumour images. In response, we propose a Dilated Groups U-Net architecture that incorporates
dilated and grouped convolution techniques to enhance the model's receptive field, significantly
improving tumour segmentation accuracy and detail. Crucially, we integrate quantum machine
learning within our transfer learning framework, achieving not only enhanced segmentation precision
but also a marked reduction in computational time. This innovative approach leverages the parallel
processing capabilities of quantum computing, accelerating the training process and potentially
enabling more efficient model adaptation to new datasets.
Our approach not only demonstrates superior performance over traditional methods but also
underscores the promising integration of quantum computing in medical image analysis, setting a
foundation for future research and clinical applications."
