An efficient framework based on large foundation model for cervical cytopathology whole slide image screening

Abstract

Cervical cytopathology whole slide image (WSI) screening primarily relies on detection-based approaches, which are limited by the high cost and labor-intensive nature of detailed annotations. Multiple Instance Learning (MIL), a weakly supervised paradigm using only slide-level labels, offers a promising alternative. However, existing MIL methods often depend on frozen pretrained models or self-supervised learning for feature extraction, which are either ineffective or computationally inefficient. To address these challenges, we propose a novel and efficient framework for cervical cytopathology WSI classification that leverages unsupervised and weakly supervised learning to enhance patch-level feature extraction. Specially, To tackle the high computational cost of training, our method introduces a mean pooling (MP)-based strategy to filter out high-risk patches, reducing the number of patches based on the sparse and dispersed nature of abnormal cells in WSIs. Additionally, we employ parameter-efficient fine-tuning (PEFT), where only the additional linear layers are trained, to significantly reduce the number of trainable parameters. Extensive experiments on the CSD and FNAC 2019 datasets demonstrate that our method consistently enhances the performance of various MIL frameworks, achieves state-of-the-art (SOTA) results, and enables faster inference speeds. Notably, on the CSD dataset, our method achieves an 8.87% improvement in specificity compared to existing approaches while maintaining the same sensitivity level of 97.84%. The code and trained models are publicly available at https://github.com/CVIU-CSU/TCT-InfoNCE.