New AI model explains its cervical cancer decisions.
A new pathology model proves that AI can explain its cervical cancer diagnoses without sacrificing speed or accuracy.
How can a doctor trust an AI that points to a tumor but cannot explain why? In cervical cytology, conventional deep learning models often act as black boxes. They spot abnormal cells but leave pathologists guessing whether the algorithm found a true biological marker or just a speck of dust on the slide.
This is where CERVEX, a new pathology-foundation-model framework, changes the conversation. It proves that explainability does not have to be an afterthought or a computational burden. By linking spatial evidence directly to its final score, it challenges the industry assumption that we must choose between a model’s accuracy and its transparency.
The math behind the map
Most models use post-hoc tools to guess why an image was flagged. CERVEX builds this directly into its architecture. The model calculates its final class score by taking the spatial mean of a class-specific evidence map and adding a learned bias.
This means the visual heatmap and the numerical score come from the exact same forward calculation. There is no translation layer to fail. It also generates a detailed research report including nuclear geometry and the nucleus-to-cytoplasm (N:C) ratio.
What the data shows
Let’s look at the performance metrics. On the RIVA dataset, CERVEX distinguished abnormal cells from normal ones with an AUROC of 0.911 (SD 0.009), a sensitivity of 0.869, and a specificity of 0.796. When tested on 101 graded abnormal slides without using cell coordinates, its fixed-grid analysis separated low-grade from high-grade disease with an AUROC of 0.823.
The model also excelled at fine-grained measurements on the SIPaKMeD dataset. It achieved a nucleus segmentation Dice score of 0.932 and a complete-cell segmentation Dice score of 0.942. Its measurements for nuclear area fraction and N:C ratio correlated with expert benchmarks at 0.977 and 0.962 respectively.
- An AUROC of 0.911 for identifying abnormal cells.
- A complete-cell segmentation Dice score of 0.942.
- A high correlation of 0.962 for the critical N:C ratio.
- An incredibly low processing overhead of just 1.9%.
The reality check
No tool is perfect, and the researchers are transparent about the limits. In a comparison using the CRIC dataset, CERVEX did not show a statistically significant accuracy difference compared to conventional classifiers. It did not outperform them in raw accuracy, but it matched them while providing full explainability.
This flat performance on CRIC is actually a crucial data point. It suggests that explainability is not a magic booster for accuracy. Instead, the real win is operational safety. Pathologists get the same level of accuracy but with a clear map of why the AI made its choice, all for a tiny 1.9% runtime overhead. This makes clinical validation a much safer bet.
Read the full study in medRxiv.
