Histopathological diagnosis of Ovine Pulmonary Adenocarcinoma (OPA) based on ensemble model.

Abstract

BACKGROUND: Ovine pulmonary adenocarcinoma (OPA) is an infectious lung tumour caused by the Jaagsiekte Sheep Retrovirus. Histological examination is the cornerstone of OPA diagnosis and provides the final morphological basis for diagnosis. However, traditional pathology faces challenges, such as complex image interpretation and reliance on subjective judgment. Ensemble learning models have been increasingly applied to medical image classification. In this study, we constructed a dataset of 69,592 images (OPA: 33,609; non-OPA: 35,983) and divided it by employing a phased dataset division strategy. After evaluating DenseNet, EfficientNet, Res2Net101, and ResNet152, Res2Net101 was selected as the best-performing base model, and ensemble learning was conducted using two strategies: output-layer fusion (Efficient-Res2Net-L) and feature fusion (Efficient-Res2Net). Model performance was evaluated using accuracy, precision, Recall, and F1 score. Anti-peeking validation was conducted using five whole-slide images (three OPA, two non-OPA) not included in the dataset. An additional 600 image blocks were used to compare performance of the model with that of pathologists. RESULTS: Res2Net101 achieved the highest accuracy (94.3%) on the test set, whereas EfficientNet made the fewest misjudgements (11) in the anti-peeking image verification. EfficientNet also outperformed others in the comparison with pathologists (accuracy: 95.0%, specificity: 91.3%, sensitivity: 98.7%). The output-layer fusion model Efficient-Res2Net-L slightly outperformed feature fusion. Efficient-Res2Net showed improved accuracy (96.5%), specificity (93.7%), and sensitivity (99.3%), surpassing the performance of junior pathologists and approaching the performance of senior pathologists, with differences reduced to 2.3% and 5%, respectively. CONCLUSION: The integrated model Efficient-Res2Ne demonstrates high accuracy and robustness. Suspicious lesion areas can be identified through rapid initial diagnosis of tissue slice images, assisting pathologists in efficiently completing the final histological diagnosis. This is a valuable tool for improving diagnostic workflow efficiency.