Automatic segmentation of epithelium in cervical whole slide images to assist diagnosis of cervical intraepithelial neoplasia

Abstract

Cervical Intraepithelial Neoplasia (CIN) is a precancerous state of the cervix, and the correct diagnosis of its level of severity (CIN grade) allows to determine patient treatment and prevent invasive carcinoma. It is diagnosed from histological samples by visually estimating the percentage of epithelial width covered by neoplastic cells, although the subjective nature of this estimation has proved to cause high levels of inter-observer variability in the diagnosis. Immunohistochemical (IHC) stains have been used to increase inter-observer agreement, and to perform immune cell analysis for cervical cancer research, but the manual quantification of IHC is still prone to subjective interpretation, and unfeasible to implement on a large scale. Digital pathology enables the digitalization of samples into Whole Slide Images (WSI), using tissue scanners that apply the principles of light microscopy. Because CIN occurs in the epithelial tissue, the first step towards a computer-assisted quantification of IHC stains would be to automatically segment the epithelium in WSIs. Deep Learning (DL) has been used in digital pathology to detect histological regions with lesions, quantify biomarkers and segment tissue types. In digital pathology, image properties such as color, brightness, contrast or blurriness may vary based on the scanner, the tissue preparation and the technician handling the equipment. The application of DL for image analysis in this field must be robust to image variability, and deliver consistent results regardless of the origin or handling of the image. Stability Training (ST) was evaluated as a potential method to increase robustness of DL models in digital pathology. During model training, ST generates a distorted version of the training images, and penalizes differences between the prediction of the original image and the prediction of the distorted version. This work evaluated if the application of ST could improve the DL models’ robustness to image variations caused by the use of different IHC stains and different scanners, and the results obtained suggest that it makes DL models better at handling WSIs with image variations. Three scanner models (NanoZoomer-HT, NanoZoomer-XR and NanoZoomer-S360) and three IHC stains (p16, CD3 and CD8) were used for this work. 114 WSIs stained with the p16 IHC and scanned with the NanoZoomer-HT (p16-HT slides) were used to train a set of DL models with different weights for the stability component (α) and different distortion combinations, including a model with no ST for reference (α=0). The models were tested on six test sets (23- 24 WSIs each) with different IHC stain and scanner combinations: p16-HT, p16-XR, p16-S360, CD3-HT and CD8-HT. An additional test set was used to evaluate robustness to blurriness caused by lack of lens focus (p16-HT-out of focus). Area Under the ROC Curve (AUC) was used to evaluate model performance. The models with ST outperformed the model without ST in all test sets. All models had a similar performance on the p16-HT test set (AUC of 0.978-0.985). For p16-HT-out of focus, the model with no ST had a significantly lower performance than the best model with ST (AUC 0.876 vs 0.939). The same was observed for p16-XR (AUC 0.913 vs 0.962), p16-S360 (AUC 0.840 vs 0.962), CD3-HT (AUC 0.973 vs 0.985) and CD8-HT (AUC 0.961 vs 0.980). Distortion combination 3 (high image distortion) and α=10 were provided the best ST results.