Clifford Convolutional Neural Networks for Lymphoblast Image Classification

This paper features convolutional neural network (CNN) models on Clifford algebras applied to a medical image classification task, namely the diagnosis of acute lymphoblastic leukemia (ALL). ALL is a type of cancer identified by malformed lymphocytes, known as lymphoblasts, in the bloodstream. The image classification task aims to discriminate healthy cells from lymphoblasts. This work shows that CNNs featuring parameters in Clifford algebras significantly outperform real-valued networks of equivalent size in this application. Indeed, the real-valued and a Clifford CNN achieved an average accuracy of 94.60% and 97.02%, respectively, in the ALL-IDB dataset with a 50% train-test split. Moreover, we present smaller versions of Clifford CNNs with roughly 75% fewer parameters that yielded a 96.50% average accuracy. The results reported in this work are comparable to high-end models in the literature despite having several orders of magnitude fewer parameters.