Uncertainty Quantification in DL Models for Cervical Cytology

We aim to investigate the correlation between a model’s predictive capability and its uncertainty estimates for cervical cytology classification through two research questions:

• How effective is uncertainty quantification for enhancing the model’s predictive capacity in high-confidence intervals?
• Are low confidence intervals an indicator of incorrect predictions?

Deep Learning (DL) has demonstrated significant promise in digital pathological applications both histopathology and cytopathology. However, the majority of these works primarily concentrate on evaluating the general performance of the models and overlook the crucial requirement for uncertainty quantification which is necessary for real-world clinical application. In this study, we examine the change in predictive performance and the identification of mispredictions through the incorporation of uncertainty estimates for DL-based Cervical cancer classification. Specifically, we evaluate the efficacy of three methods—Monte Carlo(MC) Dropout, Ensemble Method, and Test Time Augmentation(TTA) using three metrics: variance, entropy, and sample mean uncertainty. The results demonstrate that integrating uncertainty estimates improves the model’s predictive capacity in high-confidence regions, while also serving as an indicator for the model’s mispredictions in low-confidence regions.

We conduct a comparative analysis of a model’s predictive performance with and without uncertainty quantification in high-confidence intervals. Specifically, we analyse the model’s predictive performance and quantify the number of samples associated with incorrect predictions in regions of low confidence. The dataset utilized in this study is derived from the cervical cell classification collection found within the CRIC Searchable Image Database and use ResNet50 for classifying cervical cells into Normal and Abnormal categories while also estimating associated uncertainty values. We then assessed performance using AUC and Misprediction rate (100 - Accuracy)% .

Table 2 shows a 7-12% improvement in AUC across all methods in high-confidence intervals compared to the baseline AUC of 86.3 ± 2.24%. Figure 1 highlights more correct than incorrect predictions in low-uncertainty ranges, indicating that uncertainty estimates enhance predictive performance in high-confidence areas. Table 3 reveals a 10.1-19.07% increase in misprediction rates as thresholds for low-confidence intervals rise, while Figure 1 links higher uncertainty ranges to more mispredictions for MC Dropout (except Entropy). This suggests low-confidence intervals often indicate incorrect predictions.