Título: Active Deep Reinforcement Learning to Support Breast Cancer Risk Labeling
Autores: Giulia Zanon de Castro, Cássia Veiga & Frederico Gadelha Guimarães
Resumo: Breast cancer is one of the most common cancers in women worldwide. Early detection of breast cancer plays a key role in alleviating pressure on healthcare services, decreasing morbidity and mortality, and minimizing the economic burden on public health systems. While machine learning (ML) has shown promise in medical imaging, its effectiveness is limited by the scarcity of labeled data. In this context, this study explores the use of Reinforcement Learning (RL) to optimize sample selection for annotating mammographic images. Using a subset of the CBIS–DDSM dataset, which is available on Kaggle, we implemented an RL–based strategy combined with an InceptionV3 model pre–trained on RadImageNet for feature extraction. A custom environment was developed in which the RL agent learns to prioritize the most informative samples, aiming to reduce annotation costs in resource–constrained settings. Results show that the RL–based approach outperforms random sampling, achieving comparable accuracy to the fully supervised baseline (71% F–score) using fewer labeled samples. The RL model exhibited a consistent improvement trend without saturation, indicating potential for further gains with continued training. By addressing the challenge of efficient labeling, this work contributes to the development of more generalizable and cost–effective diagnostic models. The approach is particularly relevant in multi–institutional scenarios with limited and heterogeneous data, and can be further enhanced by incorporating additional clinical information to improve diagnostic precision.
Palavras-chave: machine learning; reinforcement learning; active learning; computer vision; data labeling; breast cancer; health care.
Páginas: 8
Código DOI: 10.21528/CBIC2025-1191910
Artigo em PDF: CBIC_2025_paper1191910.pdf
Arquivo BibTeX:
CBIC_2025_1191910.bib