LoRaCT: A Low Rank Adaptation-Based CNN-Transformer Model for Cervical Cancer Detection in Histopathological Images
Bhaswati Singha Deo, Mayukha Pal, Prasanta Kumar Panigrahi, Asima Pradhan · 2025
Cervical cancer is the fourth common cancer among women worldwide. The diagnosis and classification of cancer are extremely important, as it influences the optimal treatment and length of survival. Histopathological image analysis, recognized as the gold standard for cervical cancer diagnosis, is vital for its early detection. However, the varied morphological characteristics of cervical cancer make accurate manual classification challenging. Traditional diagnostic methods employed by clinicians are often time-consuming and susceptible to errors. Computer-Aided Diagnosis (CAD) systems can assist in the accurate and efficient detection of cancer in histopathological images. This study introduces an automated classification network leveraging a Low Rank Adaptation-based CNN-Transformer (LoRaCT) model. The LoRaCT model integrates convolutional neural networks (CNNs) for extracting local features with Vision Transformers (ViTs) for capturing global context. To address computational efficiency, Low Rank Adaptation (LoRa) layers are employed, significantly reducing the number of parameters while maintaining model performance. The LoRaCT model achieved an average accuracy of 95.23% accuracy on the Caishi dataset, demonstrating its potential for effective and efficient AI-driven cervical cancer detection. The LoRaCT model has achieved comparable accuracy to the standard ViT model, which used the same hyperparameters in this study, with approximately$\mathbf{9 8. 3 4 \%}$fewer parameters. This approach not only achieves high accuracy but also offers a computationally efficient solution, advancing the field of automated histopathological image analysis.