What question did this study set out to answer?

This study aims to quantify how well AI models trained on single-cell images perform when diagnosing crowded cell groups in cervical cytology.

Diagnostic performance of artificial intelligence models trained on scattered single‐cell images is not preserved for hyperchromatic crowded cell groups in cervical cytology

View Full Paper

Ask AI

Bookmark

Authors

Randomized trial quantifies AI performance loss in diagnosing hyperchromatic crowded cell groups, suggesting a need for model redesign.

This study aims to quantify how well AI models trained on single-cell images perform when diagnosing crowded cell groups in cervical cytology.
Binary convolutional neural network models were developed using a scattered cell data set of 101 cases and 1062 images.
Models trained on scattered cells were applied directly to an independent data set of hyperchromatic crowded cell groups (48 cases, 330 images).
Six model architectures were evaluated based on their area under the receiver operating characteristic curve (AUC).
Models achieved AUCs of 0.950 to 0.996 on scattered cells but declined to 0.385 to 0.683 on HCGs.
ConvNeXt-Tiny had the highest AUC of 0.683 on HCGs, significantly lower than its 0.996 on scattered cells.
All models showed significantly lower performance on HCGs compared to the scattered cell data set.

Tanaka et al. (2026) studied this question.

View Full Paper

Ask AI

Bookmark

View Full Paper

Ask AI

Bookmark

Authors

Randomized trial quantifies AI performance loss in diagnosing hyperchromatic crowded cell groups, suggesting a need for model redesign.

This study aims to quantify how well AI models trained on single-cell images perform when diagnosing crowded cell groups in cervical cytology.
Binary convolutional neural network models were developed using a scattered cell data set of 101 cases and 1062 images.
Models trained on scattered cells were applied directly to an independent data set of hyperchromatic crowded cell groups (48 cases, 330 images).
Six model architectures were evaluated based on their area under the receiver operating characteristic curve (AUC).
Models achieved AUCs of 0.950 to 0.996 on scattered cells but declined to 0.385 to 0.683 on HCGs.
ConvNeXt-Tiny had the highest AUC of 0.683 on HCGs, significantly lower than its 0.996 on scattered cells.
All models showed significantly lower performance on HCGs compared to the scattered cell data set.

Tanaka et al. (2026) studied this question.

Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context:

View Full Paper

Ask AI

Bookmark