Ensemble based in transfer learning for cytological classification in pleural fluid

Pleural effusion cytology is critical for diagnosing benign and malignant conditions, yet manual interpretation remains time-consuming and prone to subjectivity. The increasing burden of malignant pleural effusion in resource-constrained settings highlights the need for automated diagnostic solutions. This study presents an ensemble deep learning framework combining ResNet50V2, DenseNet121, and InceptionV3 architectures with transfer learning for classifying pleural cytology images into negative for malignancy (NFM) and malignant (MAL) categories. Three scenarios were evaluated: no data augmentation, 50% augmentation, and 300% augmentation. A local dataset of 1,292 images from Hospital Nacional Cayetano Heredia and an external Kaggle dataset (693 images) were used for training, validation, and independent testing. Performance was measured using accuracy, precision, recall, and F1-score across individual models and ensemble voting strategies (hard and soft voting). The ResNet + DenseNet ensemble with soft voting and 300% data augmentation achieved the highest accuracy (96.2% on the local dataset; 89.6% on the external dataset), outperforming individual models across all scenarios. Increasing the dataset through data augmentation significantly improved generalization and robustness. The proposed ensemble-based approach supports cytological diagnosis, potentially reducing diagnostic uncertainty in pleural carcinoma detection. Our findings demonstrate that ensemble deep learning models, optimized with data augmentation, can provide accurate and reproducible diagnostic support for pleural cytology, offering practical potential for deployment in low-resource healthcare settings and contributing to improved cancer diagnosis accessibility.