Few-Shot Learning for Cytological Classification of Primary Lung Cancer and Pulmonary Metastases During Endobronchial Ultrasound.
Authors
Affiliations (6)
Affiliations (6)
- Department of Mechanical Engineering, College of Engineering, National Yang Ming Chiao Tung University, Hsinchu 300, Taiwan.
- Department of Medicine, National Taiwan University Cancer Center, Taipei 106, Taiwan.
- Department of Internal Medicine, National Taiwan University Hospital, Taipei 100, Taiwan.
- Department of Internal Medicine, National Taiwan University Hsin-Chu Hospital, Hsinchu 300, Taiwan.
- Department of Laboratory Medicine, National Taiwan University Hospital, National Taiwan University College of Medicine, Taipei 100, Taiwan.
- Department of Electrical Engineering, College of Electrical Engineering and Computer Science, National Taiwan University, Taipei 106, Taiwan.
Abstract
<b>Background</b>: Endobronchial ultrasound (EBUS) is widely used for the diagnosis of pulmonary lesions. When combined with rapid on-site cytologic evaluation (ROSE), it can improve diagnostic accuracy and facilitate early identification of malignancy origin. However, differentiating primary lung cancer from metastatic tumors (e.g., breast or colorectal origin) during ROSE remains challenging due to limited cytopathology support and morphological similarities between tumor cells. This study aimed to develop a computer-aided diagnostic (CAD) system for classifying malignancy types from limited cytological samples to facilitate clinical decision-making. <b>Methods</b>: We utilized a retrospective dataset of cytological images obtained during EBUS procedures between November 2018 and June 2024. The study focused on classifying three malignancies: pulmonary adenocarcinoma, metastatic breast cancer, and metastatic colorectal cancer. A deep learning-based model (PLFCH) was developed, integrating few-shot learning, parameter-efficient fine-tuning, and a hybrid CNN-Transformer architecture to address limited annotated data. <b>Results</b>: A total of 41 patients with 346 cytological images were included. Under a 3-way 5-shot setting, the proposed model achieved an accuracy of 49.26%, outperforming existing few-shot learning methods, with precision, recall, and F1-score of 0.477, 0.493, and 0.480, respectively. Increasing the number of reference images to 20 per class further improved accuracy to 55.48%. <b>Conclusions</b>: The proposed framework demonstrated improved performance for cytological classification under limited data conditions. These findings provide an early proof of concept for applying few-shot learning to cytological assessment during EBUS procedures. Further improvements in model performance and validation in prospective, real-time clinical settings are required before its potential role in assisting diagnostic decision-making can be established.