Automated diagnosis of chronic endometritis from hysteroscopy images: a retrospective single-centre diagnostic accuracy study

Abstract Background Chronic endometritis (CE) is diagnosed from an endometrial biopsy stained for plasma-cell markers, an invasive and only moderately reproducible standard. Hysteroscopic signs are visible in the same visit but subtle, and their accuracy is disputed: meta-analyses report a summary ROC area of 0.93, whereas large single-centre series report accuracies near 0.60 to 0.70. We asked how accurately CE can be diagnosed automatically from routine hysteroscopy images. Methods We analysed 424 hysteroscopic examinations (264 CE+, 160 CE−) from one reproductive medicine centre, labelled from the routine CD38/CD138 report at one or more stromal plasma cells in the densest high-power field. Up to four pre-biopsy frames per examination were encoded with a frozen ImageNet-pretrained ConvNeXt-Tiny network, mean-pooled and classified with a radial-basis-function support vector machine. Accuracy was estimated by five-fold cross-validation grouped by woman, the operating threshold set by Youden’s J within each training fold and confidence intervals by stratified bootstrap. Results The area under the ROC curve was 0.640 (95% CI 0.581 to 0.696). Sensitivity was 0.788 (0.735 to 0.833) at a specificity of 0.463 (0.388 to 0.544), so 86 of the 160 women without histological CE were classified as positive; accuracy was 0.665. The operators’ own recorded signs gave sensitivity 0.35 and specificity 0.59. Conclusions Routine hysteroscopy images carry a reproducible but weak signal for CE, placing automated reading with the single-centre series rather than the pooled estimate. One frozen-feature model cannot identify the performance upper bound, but at this specificity the approach neither replaces nor triages histopathology before embryo transfer.