Improving Explainability of the Attention Branch Network with CAM Fostering Techniques in the Context of Histological Images

Pedro Miguel, Alessandra Lumini, Giuliano Cardozo Medalha, Guilherme Freire Roberto, Guilherme Rozendo, Adriano Cansian, Thaína Tosta, Marcelo Z. do Nascimento, Leandro Neves

2024

Abstract

Convolutional neural networks have presented significant results in histological image classification. Despite their high accuracy, their limited interpretability hinders widespread adoption. Therefore, this work proposes an improvement to the attention branch network (ABN) in order to improve its explanatory power through the gradient-weighted class activation map technique. The proposed model creates attention maps and applies the CAM fostering strategy to them, making the network focus on the most important areas of the image. Two experiments were performed to compare the proposed model with the ABN approach, considering five datasets of histological images. The evaluation process was defined via quantitative metrics such as coherency, complexity, confidence drop, and the harmonic average of those metrics (ADCC). Among the results, the proposed model through the ResNet-50 was able to provide an improvement of 4.16% in the average ADCC metric and 3.88% in the coherence metric when compared to the respective ABN model. Considering the DesneNet-201 network as the explored backbone, the proposed model achieved an improvement of 14.87% in the average ADCC metric and 9.77% in the coherence metric compared to the corresponding ABN model. The contributions of this work are important to make the results via computer-aided diagnosis more comprehensible for clinical practice.

Paper Citation

in Harvard Style

Miguel P., Lumini A., Cardozo Medalha G., Freire Roberto G., Rozendo G., Cansian A., Tosta T., Z. do Nascimento M. and Neves L. (2024). Improving Explainability of the Attention Branch Network with CAM Fostering Techniques in the Context of Histological Images. In Proceedings of the 26th International Conference on Enterprise Information Systems - Volume 1: ICEIS; ISBN 978-989-758-692-7, SciTePress, pages 456-464. DOI: 10.5220/0012595700003690

in Bibtex Style

@conference{iceis24,
author={Pedro Miguel and Alessandra Lumini and Giuliano Cardozo Medalha and Guilherme Freire Roberto and Guilherme Rozendo and Adriano Cansian and Thaína Tosta and Marcelo Z. do Nascimento and Leandro Neves},
title={Improving Explainability of the Attention Branch Network with CAM Fostering Techniques in the Context of Histological Images},
booktitle={Proceedings of the 26th International Conference on Enterprise Information Systems - Volume 1: ICEIS},
year={2024},
pages={456-464},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0012595700003690},
isbn={978-989-758-692-7},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 26th International Conference on Enterprise Information Systems - Volume 1: ICEIS
TI - Improving Explainability of the Attention Branch Network with CAM Fostering Techniques in the Context of Histological Images
SN - 978-989-758-692-7
AU - Miguel P.
AU - Lumini A.
AU - Cardozo Medalha G.
AU - Freire Roberto G.
AU - Rozendo G.
AU - Cansian A.
AU - Tosta T.
AU - Z. do Nascimento M.
AU - Neves L.
PY - 2024
SP - 456
EP - 464
DO - 10.5220/0012595700003690
PB - SciTePress