Federated Approach for Lung and Colon Cancer Classification

Deep learning is fueled by massive data. However, medical data availability is a challenge affecting the robustness of models for Computer-Aided Diagnostics. Several factors contribute to the limited amount of labeled data. One is the expertise involved in annotating biopsies and scans collected fro...

Full description

Saved in:

Bibliographic Details
Published in:	2022 19th International Computer Conference on Wavelet Active Media Technology and Information Processing (ICCWAMTIP) pp. 1 - 8
Main Authors:	Agbley, Bless Lord Y., Li, Jianping, Haq, Amin Ul, Bankas, Edem Kwedzo, Adjorlolo, Gideon, Agyemang, Isaac Osei, Ayekai, Browne Judith, Effah, Derrick, Adjeimensah, Isaac, Khan, Jalaluddin
Format:	Conference Proceeding
Language:	English
Published:	IEEE 16-12-2022
Subjects:	Biological system modeling Colon Cancer Computational modeling Data models Data privacy Deep learning Federated learning Lung Lung Cancer Training Transfer learning VGG16
Online Access:	Get full text
Tags:	Add Tag No Tags, Be the first to tag this record!

Description
Summary:	Deep learning is fueled by massive data. However, medical data availability is a challenge affecting the robustness of models for Computer-Aided Diagnostics. Several factors contribute to the limited amount of labeled data. One is the expertise involved in annotating biopsies and scans collected from laboratories. Another is the sensitive nature of medical information. This research, therefore, focuses on using data obtained on different diseases using the same technique to increase the number of data available to boost the automatic feature engineering capability of deep learning. Hence, the paper studies a multi-center-based training of a model capable of classifying two different diseases into their sub-classes. Data for each disease is hosted on separate devices, keeping the original data private to that device. VGG 16 is trained locally by each center, and the parameters are shared and aggregated for the global model. We utilized the LC25000 dataset of Lung and Cancer biopsy images for our experiment. The global model was then tested separately with client 1 (lung) and client 2 (colon) test sets. We also performed centralized learning (CL) by combining the four classes used in the decentralized experiment. Very high results were obtained by our approach, outperforming the states-of-the arts while preserving data privacy.
ISBN:	9781665493871 1665493879
ISSN:	2576-8964
DOI:	10.1109/ICCWAMTIP56608.2022.10016590