Multi-stage Fine-Tuning Deep Learning Models Improves Automatic Assessment of the Rey-Osterrieth Complex Figure Test

Schuster B, Kordon FJ, Mayr M, Seuret M, Jost S, Kessler J, Christlein V (2023)

Publication Type: Conference contribution

Publication year: 2023

Journal

Lecture Notes in Computer Science Springer Verlag

Publisher: Springer

Series: Lecture Notes in Computer Science

City/Town: Cham

Book Volume: 14187

Pages Range: 3-19

Conference Proceedings Title: Document Analysis and Recognition - ICDAR 2023

Event location: San José, CA US

ISBN: 9783031416750

DOI: 10.1007/978-3-031-41676-7_1

Abstract

The Rey-Osterrieth Complex Figure Test (ROCFT) is a widely used neuropsychological tool for assessing the presence and severity of different diseases. It involves presenting a complex illustration to the patient who is asked to copy it, followed by recall from memory after 3 and 30 min. In clinical practice, a human rater evaluates each component of the reproduction, with the overall score indicating illness severity. However, this method is both time-consuming and error-prone. Efforts have been made to automate the process, but current algorithms require large-scale private datasets of up to 20,000 illustrations. With limited data, training a deep learning model is challenging. This study addresses this challenge by developing a fine-tuning strategy with multiple stages. We show that pre-training on a large-scale sketch dataset with initialized weights from ImageNet significantly reduces the mean absolute error (MAE) compared to just training with initialized weights from ImageNet, e.g., ReXNet-200 from 3.1 to 2.2 MAE. Additionally, techniques such as stochastic weight averaging (SWA) and ensembling of different architectures can further reduce the error to an MAE of 1.97.

Authors with CRIS profile

Florian Johannes Kordon Lehrstuhl für Informatik 5 (Mustererkennung) Martin Mayr Lehrstuhl für Informatik 5 (Mustererkennung) Mathias Seuret Lehrstuhl für Informatik 5 (Mustererkennung) Vincent Christlein Lehrstuhl für Informatik 5 (Mustererkennung)

Involved external institutions

Universitätsklinikum Köln DE Germany (DE)

How to cite

APA:

Schuster, B., Kordon, F.J., Mayr, M., Seuret, M., Jost, S., Kessler, J., & Christlein, V. (2023). Multi-stage Fine-Tuning Deep Learning Models Improves Automatic Assessment of the Rey-Osterrieth Complex Figure Test. In Fink, G.A., Jain, R., Kise, K., Zanibbi, R. (Eds.), Document Analysis and Recognition - ICDAR 2023 (pp. 3-19). San José, CA, US: Cham: Springer.

MLA:

Schuster, Benjamin, et al. "Multi-stage Fine-Tuning Deep Learning Models Improves Automatic Assessment of the Rey-Osterrieth Complex Figure Test." Proceedings of the Document Analysis and Recognition - ICDAR 2023, San José, CA Ed. Fink, G.A., Jain, R., Kise, K., Zanibbi, R., Cham: Springer, 2023. 3-19.

BibTeX: Download