RESUMEN
OBJECTIVES: Histopathological diagnosis of colposcopically identified cervical lesions is a critical step for the recognition of cervical cancer precursors requiring treatment. Although there have been efforts to standardize the histologic diagnosis of cervical biopsy specimens, in terms of terminology and use of biomarkers, there is no uniform approach in the pathology community. Adjunctive p16 immunohistochemistry (IHC) can highlight precancer diagnoses, with use recommendations outlined by the Lower Anogenital Squamous Terminology project. METHODS: We assessed the diagnostic reproducibility of cervical histopathological biopsy specimens with and without p16 staining among 2 expert pathologists. RESULTS: Interpretation of p16 IHC as positive vs negative was highly reproducible (92.5% agreement, κ = 0.85); greater variation was seen in the choice of which biopsy specimens required adjunctive p16 staining (78.0% agreement, κ = 0.43). Adjunctive p16 IHC did not significantly increase diagnostic agreement under multitiered grading systems (benign vs cervical intraepithelial neoplasia [CIN] 1/low-grade squamous intraepithelial lesion vs atypical squamous metaplasia vs CIN2/high-grade squamous intraepithelial lesion [HSIL] vs CIN3/HSIL-CIN3 vs cancer) (65.5% agreement, κ = 0.56 without p16; 70.0% agreement, κ = 0.58 with p16). However, when dichotomizing diagnoses based on clinical management (less than HSIL vs HSIL+), diagnostic agreement increased with p16 IHC (90.5% agreement, κ = 0.79 without p16; 92.0% agreement, κ = 0.84 with p16). For biopsy specimens taken from women positive for human papillomavirus (HPV) type 16, agreement was similar with or without adjunctive p16 (κ = 0.80 without p16; κ = 0.78-0.80 with p16). In contrast, p16 IHC substantially improved diagnostic agreement for cervical biopsy specimens taken from women positive for other high-risk HPV strains, producing improvements in κ from 0.03 to 0.24. CONCLUSIONS: Adjunctive p16 immunostaining provides useful information in the evaluation of cervical biopsies for precancer. In our study, we have demonstrated that it is highly reproducible between 2 pathologists, although the decision of which biopsies warrant its use is less so. Furthermore, although p16 IHC showed a limited increase in diagnostic reproducibility for all biopsies included in our study, it did demonstrate a more sizable gain in biopsies negative for HPV 16 but positive for other high-risk genotypes. Further studies are needed to clarify the role of p16 IHC and how it can be optimized for the detection of cervical precancer, particularly in HPV-vaccinated populations where types other than HPV 16 are relatively more important.
RESUMEN
Background: In digital pathology, image properties such as color, brightness, contrast and blurriness may vary based on the scanner and sample preparation. Convolutional Neural Networks (CNNs) are sensitive to these variations and may underperform on images from a different domain than the one used for training. Robustness to these image property variations is required to enable the use of deep learning in clinical practice and large scale clinical research. Aims: CNN Stability Training (CST) is proposed and evaluated as a method to increase CNN robustness to scanner and Immunohistochemistry (IHC)-based image variability. Methods: CST was applied to segment epithelium in immunohistological cervical Whole Slide Images (WSIs). CST randomly distorts input tiles and factors the difference between the CNN prediction for the original and distorted inputs within the loss function. CNNs were trained using 114 p16-stained WSIs from the same scanner, and evaluated on 6 WSI test sets, each with 23 to 24 WSIs of the same tissue but different scanner/IHC combinations. Relative robustness (rAUC) was measured as the difference between the AUC on the training domain test set (i.e., baseline test set) and the remaining test sets. Results: Across all test sets, The AUC of CST models outperformed "No CST" models (AUC: 0.940-0.989 vs. 0.905-0.986, p < 1e - 8), and obtained an improved robustness (rAUC: [-0.038, -0.003] vs. [-0.081, -0.002]). At a WSI level, CST models showed an increase in performance in 124 of the 142 WSIs. CST models also outperformed models trained with random on-the-fly data augmentation (DA) in all test sets ([0.002, 0.021], p < 1e-6). Conclusion: CST offers a path to improve CNN performance without the need for more data and allows customizing distortions to specific use cases. A python implementation of CST is publicly available at https://github.com/TIGACenter/CST_v1.