|
Enhancing the Interpretability of Cervical Cancer Diagnosis: Refining Geometry-Based Heatmaps Using Ricci Flow on the SIPaKMeD Dataset
|
F Abbasi Varaki *    , A Ariyaei Motahar , M Mohammadian Amiri |
| 1.School of Medicine, Iran University of Medical Sciences, Tehran, I.R.Iran. , farhanabbasip7@gmail.com |
|
|
Abstract: (20 Views) |
Background and Objective: Cervical cancer detection is a critical task where accurate and interpretable diagnostic tools can significantly enhance clinical outcomes. Existing explainability methods, such as Grad-CAM, provide visual insights into model predictions but often lack precision in highlighting diagnostically relevant regions. To address this, we integrated Ricci Flow into the gradient computation process, aiming to refine the geometric structure of heatmaps and improve their interpretability.
Methods: In this methodological study, The SIPaKMeD dataset, comprising five categories of cervical smear images (4049 images), was preprocessed and augmented to address the class imbalance and ensure uniformity. A ResNet50 model was trained on the dataset for 40 epochs using a balanced set of whole-slide images. Ricci Flow was applied to the gradient matrix, interpreted as a Riemannian tensor, to iteratively smooth and refine the geometry of the gradient space. Heatmap quality was evaluated through supervised comparison with cropped cell images and by computing insertion and deletion metrics to quantify the alignment of heatmaps with diagnostically critical regions.
Findings: The Ricci Flow-enhanced Grad-CAM outperformed the standard Grad-CAM, achieving an insertion AUC of 0.671 and a deletion AUC of 0.153. The refined heatmaps consistently demonstrated a sharper focus on diagnostically important regions, with over 90 percent alignment with expert annotations. Additionally, the Ricci Flow-based method provided a deeper geometric understanding of the feature space, emphasizing the region’s most critical for classification.
Conclusion: The integration of Ricci Flow into Grad-CAM enhances heatmap interpretability by leveraging geometric smoothing, offering a novel approach to explainable AI in cervical cancer detection. This method not only improves model precision but also aligns with the hypothesis that neural networks transform data topology during training.
|
|
| Keywords: Ricci Flow, Explainable AI, Uterine Cervical Neoplasms, Heatmap Interpretability, SIPaKMeD Dataset, Deep Learning. |
|
|
|
Type of Study: Applicable |
Subject:
Obstetrics and Gynecology
Received: 2024/12/8 | Accepted: 2025/02/4
|
|
|
|
|
|
|
| Send email to the article author |
|
|
