A reinforcement learning approach for lipid droplet detection utilizing transfer learning and visualization of distribution patterns
Year: 2026
Type: article
Abstract: Motivation Accurate quantitative analysis of lipid droplet distribution in liver pathology images is crucial for diagnosing metabolic diseases. We previously improved detection using reinforcement learning (RL) and random forests (LiDRL v2), but RL posed challenges with training times and false positives. This study introduces LiDRL v3, which incorporates Q-table transfer learning and an abstaining strategy in stochastic random forests, achieving more efficient and accurate lipid droplet detection from limited data. Results Q-Table transfer learning accelerated RL-based detection, while the abstaining strategy reduced false positives. Kernel density estimation, spatial entropy, and shape scores were applied using local structure analysis to visually and quantitatively assess the structural features of lipid droplets overlaid on histological images. The results were qualitatively consistent with observations from two pathologists, confirming that this approach enables rapid and interpretable quantification of lipid droplet distribution, suggesting a potential application for disease classification. Availability and implementation LiDRL v3 framework was implemented in Python (version 3.10.12) and can be implemented on a CPU environment, facilitating deployment in resource-limited settings. The source code for lipid droplet detection and distribution quantification is publicly available at Zenodo (https://doi.org/10.5281/zenodo.20641137). Ethical restrictions and patient privacy considerations limit public distribution of the histopathological image data.
Source: Bioinformatics Advances
Institutions: Nippon Bunri University, Oita University
Language: English
Topic: AI in cancer detection
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,726
OpenAlex ID: W7236343632
