Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Lightweight, Interpretable Deep Learning System for Automated Detection of Cervical Adenocarcinoma In Situ (AIS)

Gabriela Fernandes

TL;DR

Problem: AIS diagnosis on histopathology is challenging but early detection is critical. Approach: a lightweight, interpretable AI pipeline combines Macenko stain normalization, patch-based extraction, EfficientNet-B3, focal loss, and Grad-CAM explanations, deployed via a Gradio virtual assistant. Findings: on CAISHI, it achieves an overall accuracy of 0.7323 with Abnormal F1 of 0.75 and Normal F1 of 0.71, including a substantial reduction in false negatives (from 86 to 46). Significance: demonstrates the feasibility of deployable, explainable AI for cervical gland pathology with potential to augment screening, education, and resource-limited workflows; limitations include dataset size and need for external validation.

Abstract

Cervical adenocarcinoma in situ (AIS) is a critical premalignant lesion whose accurate histopathological diagnosis is challenging. Early detection is essential to prevent progression to invasive cervical adenocarcinoma. In this study, we developed a deep learning-based virtual pathology assistant capable of distinguishing AIS from normal cervical gland histology using the CAISHI dataset, which contains 2240 expert-labeled H&E images (1010 normal and 1230 AIS). All images underwent Macenko stain normalization and patch-based preprocessing to enhance morphological feature representation. An EfficientNet-B3 convolutional neural network was trained using class-balanced sampling and focal loss to address dataset imbalance and emphasize difficult examples. The final model achieved an overall accuracy of 0.7323, with an F1-score of 0.75 for the Abnormal class and 0.71 for the Normal class. Grad-CAM heatmaps demonstrated biologically interpretable activation patterns, highlighting nuclear atypia and glandular crowding consistent with AIS morphology. The trained model was deployed in a Gradio-based virtual diagnostic assistant. These findings demonstrate the feasibility of lightweight, interpretable AI systems for cervical gland pathology, with potential applications in screening workflows, education, and low-resource settings.

A Lightweight, Interpretable Deep Learning System for Automated Detection of Cervical Adenocarcinoma In Situ (AIS)

TL;DR

Problem: AIS diagnosis on histopathology is challenging but early detection is critical. Approach: a lightweight, interpretable AI pipeline combines Macenko stain normalization, patch-based extraction, EfficientNet-B3, focal loss, and Grad-CAM explanations, deployed via a Gradio virtual assistant. Findings: on CAISHI, it achieves an overall accuracy of 0.7323 with Abnormal F1 of 0.75 and Normal F1 of 0.71, including a substantial reduction in false negatives (from 86 to 46). Significance: demonstrates the feasibility of deployable, explainable AI for cervical gland pathology with potential to augment screening, education, and resource-limited workflows; limitations include dataset size and need for external validation.

Abstract

Cervical adenocarcinoma in situ (AIS) is a critical premalignant lesion whose accurate histopathological diagnosis is challenging. Early detection is essential to prevent progression to invasive cervical adenocarcinoma. In this study, we developed a deep learning-based virtual pathology assistant capable of distinguishing AIS from normal cervical gland histology using the CAISHI dataset, which contains 2240 expert-labeled H&E images (1010 normal and 1230 AIS). All images underwent Macenko stain normalization and patch-based preprocessing to enhance morphological feature representation. An EfficientNet-B3 convolutional neural network was trained using class-balanced sampling and focal loss to address dataset imbalance and emphasize difficult examples. The final model achieved an overall accuracy of 0.7323, with an F1-score of 0.75 for the Abnormal class and 0.71 for the Normal class. Grad-CAM heatmaps demonstrated biologically interpretable activation patterns, highlighting nuclear atypia and glandular crowding consistent with AIS morphology. The trained model was deployed in a Gradio-based virtual diagnostic assistant. These findings demonstrate the feasibility of lightweight, interpretable AI systems for cervical gland pathology, with potential applications in screening workflows, education, and low-resource settings.

Paper Structure

This paper contains 11 sections, 3 figures.

Table of Contents

  1. Introduction
  2. Methods
  3. Dataset and Preprocessing
  4. Patch Extraction and Augmentation
  5. Model Architecture and Training
  6. Results
  7. Diagnostic Performance and Confusion Matrix Analysis
  8. Performance Across Prediction Thresholds
  9. Discussion
  10. Limitations
  11. Conclusion

Figures (3)

  • Figure 1: Confusion Matrix (after transfer learning). The final model achieved 123 True Negatives (correctly classified Abnormal) and 104 True Positives (correctly classified Normal) [65]. Crucially, the number of False Negatives (Actual Abnormal, Predicted Normal) was 46, representing a significant reduction compared to the previous model iteration [66].
  • Figure 2: Sample False Negative Images (Actual: Abnormal, Predicted: Normal). Examples of cervical gland histology images where the model incorrectly predicted 'Normal' [70]. These cases represent subtle or ambiguous presentations of AIS, highlighting the complexity of classifying real-world tissue patterns [71].
  • Figure 3: Precision, Recall, and F1-score vs. Prediction Threshold. This plot illustrates the performance metrics as a function of the classification threshold [74]. The intersection of the Abnormal Recall (orange) and Normal Recall (dashed purple with 'x') curves near a threshold of 0.45 indicates a balanced operating point for the two classes [75].