Contrastive learning-guided multiple instance learning for automated diagnosis of chronic endometritis from digitized histopathology

Chronic endometritis (CE) is a common yet underdiagnosed endometrial inflammatory disorder. This study aimed to develop and validate a deep learning model for direct CE detection from H&E-stained slides. A retrospective dataset of 757 infertile women (257 with CE and 500 without CE) was used for model training and internal validation. External validation was conducted using a cohort of 120 patients (40 CE and 80 non-CE). CE was confirmed if five or more CD138 + plasma cells were observed in at least one of ten high-power fields. The model employed a contrastive learning–based multiple instance learning architecture with a ResNet backbone and attention mechanism. Model performance was evaluated through five-fold cross-validation and external testing. The proposed model achieved an average diagnostic accuracy of 69.5% ± 2.3% and an AUC of 0.66 in internal cross-validation. In the external test cohort, accuracy remained comparable at 70.3%, with an AUC of 0.69 and a recall of 49.09%. PathSENet yielded the numerically highest AUC among the models in both internal and external evaluations. The proposed deep learning model demonstrates the feasibility of automated CE classification directly from H&E-stained whole-slide images. However, given its modest discriminative performance and limited recall, it should currently be considered a proof-of-concept auxiliary screening tool rather than a standalone alternative to CD138 immunohistochemistry. Further model optimization and multicenter validation are required before clinical implementation.