Modification of Convolutional Neural Networks for Brain Tumor Segmentation on MRI with Limited Computational Resources

Brain tumor segmentation on magnetic resonance imaging (MRI) plays a critical role in diagnosis and treatment planning but demands significant computational resources. This study proposes an approach to modify convolutional neural networks (CNNs) for tumor segmentation on hardware with limited capabilities, specifically a device with 12 GB of GPU memory and 32 GB of RAM. Using data from open sources, includingmeningiomas, glioblastomas, and metastases, a Dice Score ranging from 0.88 to 0.91 was achieved. To enhance efficiency, modifications to the U-Net architecture and adaptive image preprocessing were applied. The U-Net modification involves reducing the number of filters from 64 to 32 in the initial layers, while adaptive preprocessing employs a Gaussian Blur filter with a parameter <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\sigma$</tex> adjusted based on noise levels <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(\sigma=1.2$</tex> for SNR ≥ 20 <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{dB}, \sigma=1.5$</tex> for SNR < 20 dB). These changes, combined with mixed-precision FP16 and a resolution of 320×320, enable the model to operate effectively under resource constraints, offering a scalable solution for automated tumor analysis. The ability to function on low-end hardware is crucial for small clinics and research labs lacking powerful computational clusters, thereby broadening access to automated MRI analysis and accelerating diagnosis. For more rigorous validation, testing on the BraTS dataset or data from the Federal Center of Neurosurgery is planned.