A Comparative Study of Sparse and Dense CNNs under Pixel Threshold-Induced Sparsity for Brain Tumor Classification

This paper compares dense and sparse convolutional neural networks (CNNs) for brain tumor classification from images of 2D MRI scans, using the Minkowski Engine sparse tensor framework. Sparsity is induced by thresholding pixel intensities, retaining only pixels above a cutoff value τ. Two architectures are evaluated , a lightweight ResNet-style model and ResNet-101, in both dense and sparse variants, on a four-class dataset covering glioma, meningioma, pituitary, and no-tumor cases. Results show that threshold-induced sparsity matches or improves on dense baselines. This work originated as a mentorship project through the NAACP’s Afro-Academic, Cultural, Technological and Scientific Olympics (ACT-SO) in collaboration with Fermi National Accelerator Laboratory and Argonne National Laboratory.