Evolution of Convolutional Neural Networks for Lymphoma Classification.

We investigated using deep learning and evolutionary algorithms to classify Lymphoma, achieving a 10-fold cross-validation accuracy of 95.64% and a single run accuracy of 98.41%, surpassing average human pathologist performance.

Published on Mar 01, 2021

Arxiv Link: https://link.springer.com/chapter/10.1007/978-3-030-71051-4_1

Github

Technology

Hugo
Python 3
Openslide

Reading time: 1 minutes.

Abstract :

Lymphoma comprises over 60 subtypes, typically requiring specialized pathologists for accurate diagnosis. Our study aimed to evaluate the efficacy of Artificial Neural Networks (ANNs) and Deep Learning in Lymphoma classification, while also assessing the optimization potential of Evolutionary Algorithms (EAs). Utilizing TensorFlow and Keras for network construction, we developed an innovative framework to evolve network weights. The optimal model, a Convolutional Neural Network (CNN), achieved a tenfold cross-validation test accuracy of 95.64% and a peak single run accuracy of 98.41%. These results indicate that ANNs can surpass the diagnostic accuracy of average human pathologists. Additionally, the consistent accuracy improvements via EAs underscore their utility in enhancing ANNs for Lymphoma classification.