A.I. Strategies for Predicting Regression Potential in H&E images of Cervical Precancerous Lesions

In Finland, over 2000 women undergo Cervical intraepithelial neoplasia (CIN) treatment annually [1]. The median age of women at treatment of a CIN-lesion is approximately 30, a common age for first delivery in Finland. The treatment of a CIN-lesion has been associated with 2-3 times increased risk of preterm birth [2]. Given the unpredictable nature of substantial number of CIN2 cases spontaneously regressing to normal, many women undergo an unnecessary CIN-treatment [3]. The treatment decision of CIN patients heavily relies on the precise analysis of cell and tissue structures within histopathological images, however, the traditional diagnostic methods of manual interpretation are prone to subjectivity and human error, leading to uncertainty in treatment decisions [5]. In recent years, digital pathology has emerged as a powerful tool for cancer diagnosis and prognosis [6]. While digital pathology has shown immense promise, its full potential is yet to be realized. Deep learning (DL) based computer vision methods have already demonstrated their potential in image analysis, and
when applied to digital pathology, they offer the opportunity to unlock a level of precision that surpasses human capabilities [7]. We have developed advanced DL models for cell and tissue segmentation in digitized whole slide images (WSI), the cornerstone of digital pathology. Coupled with our novel spatial analysis techniques, we can extract histologically meaningful and human-readable features for pathologists. Our aim is to use these tools to quantify different spatially compartmentalized histomorphological features within cervical tissue to predict the regression potential of CIN2 lesions.