A Histopathologic Image Analysis for the Classification of Endocervical Adenocarcinoma Silva Patterns Depend on Weakly Supervised Deep Learning

Twenty-five percent of cervical cancers are classified as endocervical adenocarcinomas (EACs), which comprise a highly heterogeneous group of tumors. A histopathologic risk stratification system known as the Silva pattern system was developed based on morphology. However, accurately classifying such patterns can be challenging. The study objective was to develop a deep learning pipeline (Silva3-AI) that automatically analyzes whole slide image–based histopathologic images and identifies Silva patterns with high accuracy. Initially, a total of 202 patients with EACs and histopathologic slides were obtained from Qilu Hospital of Shandong University for developing and internally testing the Silva3-AI model. Subsequently, an additional 161 patients and slides were collected from seven other medical centers for independent testing. The Silva3-AI model was developed using a vision transformer and recurrent neural network architecture, utilizing multi-magnification patches, and its performance was evaluated based on a class-specific area under the receiver-operating characteristic curve. Silva3-AI achieved a class-specific area under the receiver-operating characteristic curve of 0.947 for Silva A, 0.908 for Silva B, and 0.947 for Silva C on the independent test set. Notably, the performance of Silva3-AI was consistent with that of professional pathologists with 10 years' diagnostic experience. Furthermore, the visualization of prediction heatmaps facilitated the identification of tumor microenvironment heterogeneity, which is known to contribute to variations in Silva patterns. Twenty-five percent of cervical cancers are classified as endocervical adenocarcinomas (EACs), which comprise a highly heterogeneous group of tumors. A histopathologic risk stratification system known as the Silva pattern system was developed based on morphology. However, accurately classifying such patterns can be challenging. The study objective was to develop a deep learning pipeline (Silva3-AI) that automatically analyzes whole slide image–based histopathologic images and identifies Silva patterns with high accuracy. Initially, a total of 202 patients with EACs and histopathologic slides were obtained from Qilu Hospital of Shandong University for developing and internally testing the Silva3-AI model. Subsequently, an additional 161 patients and slides were collected from seven other medical centers for independent testing. The Silva3-AI model was developed using a vision transformer and recurrent neural network architecture, utilizing multi-magnification patches, and its performance was evaluated based on a class-specific area under the receiver-operating characteristic curve. Silva3-AI achieved a class-specific area under the receiver-operating characteristic curve of 0.947 for Silva A, 0.908 for Silva B, and 0.947 for Silva C on the independent test set. Notably, the performance of Silva3-AI was consistent with that of professional pathologists with 10 years' diagnostic experience. Furthermore, the visualization of prediction heatmaps facilitated the identification of tumor microenvironment heterogeneity, which is known to contribute to variations in Silva patterns. In 2020, cervical cancer ranked as the fourth most prevalent cancer among women, with a projected 604,127 new cases and 341,831 fatalities worldwide.1Sung H. Ferlay J. Siegel R.L. Laversanne M. Soerjomataram I. Jemal A. Bray F. Global Cancer Statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries.CA Cancer J Clin. 2021; 71: 209-249Crossref PubMed Scopus (53173) Google Scholar The majority of cervical cancer cases are causally linked to high-risk human papillomavirus (HR-HPV) infection. As cervical cancer screening and HPV vaccination rates have increased, the proportion of squamous cell carcinomas has declined, while the proportion of endocervical adenocarcinomas (EACs) has risen. EACs account for approximately 25% of cervical cancers and represent a more heterogeneous group of tumors.2Cohen P.A. Jhingran A. Oaknin A. Denny L. Cervical cancer.Lancet. 2019; 393: 169-182Abstract Full Text Full Text PDF PubMed Scopus (1336) Google Scholar The management of EACs is primarily determined by the stage of the disease upon diagnosis. However, the existing staging systems for EACs, established by the International Federation of Gynecology and Obstetrics and the American Joint Committee on Cancer, do not provide precise guidelines for measuring the invasive histologic component.3Fujiwara H. Yokota H. Monk B. Treilleux I. Devouassoux-Shisheboran M. Davis A. Kim J.-W. Mahner S. Stany M. Pignata S. Ray-Coquard I. Fujiwara K. Gynecologic Cancer InterGroup (GCIG) consensus review for cervical adenocarcinoma.Int J Gynecol Cancer. 2014; 24: S96-S101Crossref PubMed Scopus (44) Google Scholar This is particularly important because different histologic subtypes of EACs have distinct molecular underpinnings that could greatly affect future therapeutic interventions.4Stolnicu S. Hoang L. Chiu D. Hanko-Bauer O. Terinte C. Pesci A. Aviel-Ronen S. Kiyokawa T. Alvarado-Cabrero I. Oliva E. Park K.J. Abu-Rustum N.R. Soslow R.A. Clinical outcomes of HPV-associated and unassociated endocervical adenocarcinomas categorized by the International Endocervical Adenocarcinoma Criteria and Classification (IECC).Am J Surg Pathol. 2019; 43: 466-474Crossref PubMed Scopus (76) Google Scholar The aim of the 2018 International Endocervical Adenocarcinoma Criteria and Classification was to stratify EACs based on HPV status and the morphology of tumor stromal invasion.5Stolnicu S. Barsan I. Hoang L. Patel P. Terinte C. Pesci A. Aviel-Ronen S. Kiyokawa T. Alvarado-Cabrero I. Pike M.C. Oliva E. Park K.J. Soslow R.A. International Endocervical Adenocarcinoma Criteria and Classification (IECC): a new pathogenetic classification for invasive adenocarcinomas of the endocervix.Am J Surg Pathol. 2018; 42: 214-226Crossref PubMed Scopus (244) Google Scholar In 2013, the Silva pattern classification system was introduced, which could predict the risk of lymph node metastasis and recurrences in HPV-associated EACs.6Roma A.A. Diaz De Vivar A. Park K.J. Alvarado-Cabrero I. Rasty G. Chanona-Vilchis J.G. Mikami Y. Hong S.R. Teramoto N. Ali-Fehmi R. Rutgers J.K.L. Barbuto D. Silva E.G. Invasive endocervical adenocarcinoma: a new pattern-based classification system with important clinical significance.Am J Surg Pathol. 2015; 39: 667-672Crossref PubMed Scopus (85) Google Scholar However, accurately classifying such patterns can be challenging due to the qualitative criteria used, which tend to induce interobserver variability among pathologists. This issue has been reported in several studies.7Paquette C. Jeffus S.K. Quick C.M. Conaway M.R. Stoler M.H. Atkins K.A. Interobserver variability in the application of a proposed histologic subclassification of endocervical adenocarcinoma.Am J Surg Pathol. 2015; 39: 93-100Crossref PubMed Scopus (38) Google Scholar, 8Parra-Herran C. Taljaard M. Djordjevic B. Reyes M.C. Schwartz L. Schoolmeester J.K. Lastra R.R. Quick C.M. Laury A. Rasty G. Nucci M.R. Howitt B.E. Pattern-based classification of invasive endocervical adenocarcinoma, depth of invasion measurement and distinction from adenocarcinoma in situ: interobserver variation among gynecologic pathologists.Mod Pathol. 2016; 29: 879-892Abstract Full Text Full Text PDF PubMed Scopus (43) Google Scholar, 9Rutgers J.K.L. Roma A.A. Park K.J. Zaino R.J. Johnson A. Alvarado I. Daya D. Rasty G. Longacre T.A. Ronnett B.M. Silva E.G. Pattern classification of endocervical adenocarcinoma: reproducibility and review of criteria.Mod Pathol. 2016; 29: 1083-1094Abstract Full Text Full Text PDF PubMed Scopus (42) Google Scholar In recent years, digital pathology has emerged as a promising new standard of care, whereby glass slides are transformed into whole slide images (WSIs) using digital slide scanners. The ability to digitize WSIs of tissue has paved the way for the introduction of artificial intelligence techniques in computational pathology, which facilitate the extraction of subvisual morphometric data. Deep learning, as a significant subfield of artificial intelligence, has shown considerable potential in the evaluation of tumor histopathology,10Srinidhi C.L. Ciga O. Martel A.L. Deep neural network models for computational histopathology: a survey.Med Image Anal. 2021; 67101813Abstract Full Text Full Text PDF PubMed Scopus (355) Google Scholar quantification of histopathology elements,11Li X. Tang Q. Yu J. Wang Y. Shi Z. Microvascularity detection and quantification in glioma: a novel deep-learning-based framework.Lab Invest. 2019; 99: 1515-1526Abstract Full Text Full Text PDF PubMed Scopus (14) Google Scholar and histology classification or subclassification.12Hekler A. Utikal J.S. Enk A.H. Berking C. Klode J. Schadendorf D. Jansen P. Franklin C. Holland-Letz T. Krahl D. von Kalle C. Fröhling S. Brinker T.J. Pathologist-level classification of histopathological melanoma images with deep neural networks.Eur J Cancer. 2019; 115: 79-83Abstract Full Text Full Text PDF PubMed Scopus (137) Google Scholar,13Ambrosini P. Hollemans E. Kweldam C.F. Leenders G.J.L.H. Stallinga S. Vos F. Automated detection of cribriform growth patterns in prostate histology images.Sci Rep. 2020; 1014904Crossref Scopus (15) Google Scholar The emergence of computational pathology has facilitated computer-assisted diagnostics and enabled a digital workflow for pathologists.14Fuchs T.J. Wild P.J. Moch H. Buhmann J.M. Computational pathology analysis of tissue microarrays predicts survival of renal clear cell carcinoma patients.Med Image Comput Comput Assist Interv. 2008; 11: 1-8PubMed Google Scholar, 15Fuchs T.J. Buhmann J.M. Computational pathology: challenges and promises for tissue analysis.Comput Med Imaging Graph. 2011; 35: 515-530Crossref PubMed Scopus (184) Google Scholar, 16Louis D.N. Feldman M. Carter A.B. Dighe A.S. Pfeifer J.D. Bry L. Almeida J.S. Saltz J. Braun J. Tomaszewski J.E. Gilbertson J.R. Sinard J.H. Gerber G.K. Galli S.J. Golden J.A. Becich M.J. Computational pathology: a path ahead.Arch Pathol Lab Med. 2016; 140: 41-50Crossref PubMed Scopus (93) Google Scholar The development of diagnostic decision support tools can greatly enhance the efficiency and accuracy of pathologists. The objective of the current study was to devise a deep learning pipeline named Silva3-AI, which can automatically scrutinize histopathologic images based on WSIs and accurately detect the Silva pattern. The efficacy of Silva3-AI was evaluated through comprehensive verification experiments conducted on independent tests and compared with the assessments of two pathologists. The hypothesis that Silva3-AI can accurately identify the Silva pattern histologically, while exhibiting good generalization ability, was tested. To accurately determine the growth pattern of a tumor, it is often necessary to conduct a microscopic examination of the entire tumor. For the purposes of this study, patients who had undergone a hysterectomy between June 2016 and June 2021were selected. The Silva pattern was deemed applicable only to cases of HPV-associated EAC. To confirm HPV status, patients were initially reviewed through electronic health records or, when these records were unavailable, confirmation was based on the presence of HPV-associated features such as koilocytosis and immunohistochemistry of p16 protein. Patients with non–HPV-associated tumors were excluded from the study. Hematoxylin and eosin (H&E)–stained slides were obtained from the Department of Pathology at Qilu Hospital of Shandong University (QLHSDU). Before scanning of the WSIs, these Silva pattern annotations were confirmed by a team of four professional gynecologic pathologists after thorough examination of H&E–stained tissue sample slides of tumor samples; these pathologists had 15 to 30 years of experience. First, each EAC sample was re-evaluated by two professional gynecologic pathologists [Xi. Zhang (15 years of experience) and Chunyan Hao (20 years of experience; Department of Pathology, School of Basic Medical Sciences and Qilu Hospital, Shandong University)] regarding the Silva pattern. The Silva pattern was assessed based on the 2018 International Endocervical Adenocarcinoma Criteria and Classification and the Endocervical Adenocarcinoma Working Group of the International Society of Gynecological Pathologists reference guide (Figure 1 and Table 1). If the annotation of Xi. Zhang and Chunyan Hao was in agreement, the patient's Silva pattern annotation was then confirmed; if the annotation of Xi. Zhang and Chunyan Hao was in disagreement, then together with the two other professional gynecologic pathologists [L.L. (with 23 years of experience; Department of Pathology, School of Basic Medical Sciences and Qilu Hospital, Shandong University) and Tingguo Zhang (with 30 years of experience; Department of Pathology, School of Basic Medical Sciences and Qilu Hospital, Shandong University)], the four pathologists made the annotation, conducted a professional, rigorous discussion, and then made a final decision on the Silva pattern. Tumors exhibiting alternative growth patterns, including gastric type, clear cell, mesonephric, and endometrioid, were excluded from the study. One representative slide was chosen for each patient, resulting in a QLHSDU data set comprising 202 patients and 202 H&E–stained slides.Table 1Silva Pattern Classification Criteria of HPV-Associated EACsSilva pattern ASilva pattern BSilva pattern C(1) Well-demarcated glands with rounded contours, frequently forming groups(2) No destructive stromal invasion(3) No single cells or cell detachment(4) Complex intraglandular growth acceptable (cribriform or papillae)(5) Lack of solid growth(6) Depth of tumor or relationship to large vessels not relevant(7) No lymphovascular invasion(1) Localized (limited, early) destructive stromal invasion arising from pattern A glands(2) Individual or small groups of tumor cells, separated from the rounded gland, often in focally desmoplastic or inflamed stroma(3) Foci may be single, multiple, or linear at base of tumor(4) Lymphovascular invasion (±)(1) Diffuse destructive stromal invasion, characterized by diffusely infiltrative glands with extensive desmoplastic response(2) Confluent growth filling a ×40 field (5 mm) glands, papillae (stroma only within papillae) or mucin lakes(3) Glands often angulated or with canalicular pattern with interspersed open glands(4) Solid, poorly differentiated component (architecturally high grade); nuclear grade is disregarded(5) Lymphovascular invasion (±)EAC, endocervical adenocarcinoma; HPV, human papillomavirus. Open table in a new tab EAC, endocervical adenocarcinoma; HPV, human papillomavirus. To assess the classifier's generalization ability, additional validations were performed on seven distinct external cohorts: Binzhou Medical University Hospital, Affiliated Hospital of Jining Medical University, Liaocheng People's Hospital, The Second Hospital of Shandong University, Shandong Provincial Hospital, Weihai Municipal Hospital, and KingMed Diagnostics. Each patient selected a representative slide, resulting in a total of 161 patients and 161 H&E–stained slides within the independent external cohort. These slides were generated at their respective institutions and subsequently scanned on whole slide scanners at QLHSDU. Notably, these external slides were not used during the training phase but rather solely during testing to evaluate the classifier's capacity to generalize across diverse sources of technical variability in glass slide preparation (Figure 2). Data collection and analysis for this research project were anonymized to protect patients' privacy and were approved by the Ethics Committee of QLHSDU (approval no. KYLL-202207-003). The slides underwent scanning using three different scanners: a WS-10 scanner (WISLEAP, Changzhou, China), at 40× equivalent magnification (0.23 μm per pixel) and digitized into ndpi format; a PANNORAMIC MIDI Ⅱ scanner (3DHISTECH, Budapest, Hungary), at 20× equivalent magnification (0.23 μm per pixel) and digitized into mrxs format; and an isanCoreroAu scanner (Roche Diagnostics, Basel, Switzerland), at 20× equivalent magnification (0.24 μm per pixel) and digitized into bif format. All pathology slides were consistent at the pixel-wise level (Supplemental Table S1). The digital slides were then subjected to manual inspection for the presence of blur artifacts, and any affected slides were rescanned before downstream data processing. The WSIs were traversed and cropped into nonoverlapping patches measuring 1024 × 1024 pixels. This was done using a sliding window approach with a stride of 1024. Patches that showed image pixel variation below 500, indicating the absence of lesion tissues, were eliminated to improve computational efficiency. The remaining patches underwent color normalization using the MacenkoNormalizer technique,17Macenko M. Niethammer M. Marron J.S. Borland D. Thomas N.E. A method for normalizing histology slides for quantitative analysis.in: IEEE International Symposium on Biomedical Imaging. From Nano to Macro, Boston, MA2009: 1107-1110Crossref Scopus (870) Google Scholar which was done to enhance convergence during the training process (Figure 3). The study used a two-step framework. The pretrained patch embedder of vision transformer–recurrent neural network (RNN) used the state-of-the-art vision transformer for feature extraction. It calculated scores for patches and ranked them based on their probability of being each Silva pattern. The predicted scores for patches within a WSI were then aggregated, resulting in the creation of semantically rich patch-level feature representations. These representations were integrated into an RNN to consolidate information across the input patches, ultimately leading to the final classification (Figure 4). To facilitate the training and optimization of networks in Silva3-AI, a random division of slides at the slide level was used, divided into distinct training and testing sets. The training process was repeated for various magnification levels, with magnifications ranging from 20× to 400× (slides were scanned with an eyepiece of 10× and multi-magnification objectives of 2×, 5×, 10×, 20×, and 40×, and the magnification levels were 20×, 50×, 100×, 200×, and 400×, respectively). This magnification level strategy is influenced by the diagnostic guidelines of pathologists, who diagnose Silva C based on low magnification of tissue, followed by the diagnosis of Silva A and B at high magnification. In Silva3-AI, Silva C was initially identified at low magnification: if the patch was identified as Silva C, its probability of being positive would approach 1; if the patch was identified as Silva A or B at low magnification, the probability of being negative would approach 0. Secondly, the patch identified as Silva A and B at low magnification would identify at high magnification again: if the patch was identified as Silva B, its probability of being positive would approach 1; while if the patch was identified as Silva A, the probability of being negative would approach 0. The trained networks underwent testing on patches extracted at the same magnification from both the QLHSDU internal test and independent external cohorts for each magnification. An index K (k = 10, 50, 100, 500, 1000, 1500, 2000) was then obtained for each patch within the slide, indicating the Top-K highest probability of being "0 or 1." These Top-K–ranked patches were subsequently used as input for an RNN, which was utilized for final classification. The last fully connected layer of the RNN was replaced with a Softmax layer to generate output of a two-dimension vector. The code used in this article is freely available (https://github.com/huggingface/pytorch-image-models/tree/main/timm, last accessed December 12, 2023). Initially, each patch underwent down-sampling to the standard 384 × 384 pixels. Before being fed into the vision transformer, the network was initialized with default weights transferred from the vit_base_patch_384 data set. The rescaling of pixels from a range of 0 to 384 to a range of 0 to 1 was achieved through division by 384, followed by Z score normalization using a mean of (0.5, 0.5, 0.5) and a standard deviation of (0.1, 0.1, 0.1). All layers were frozen except for the last fully connected layer, which was trained by using QLHSDU data. In the second stage, the frozen layers were unfrozen, and the entire network was fine-tuned to achieve optimal performance for the target task. The trainable layer parameters were optimized by referencing the cross-entropy between predictions and ground truths. The initial learning rate and weight decay were set to 1 × 10–4 and 1× 10–4, respectively; the optimizer was Adam.18Kingma D.P. Ba J. Adam: a method for stochastic optimization.arXiv. 2014; ([Preprint])https://doi.org/10.48550/arXiv.1412.6980Google Scholar A grid search was conducted for each experiment to optimize the learning rate and mini-batch size, with a learning rate decay of 0.7 every five epochs. During training, a dropout rate of 0.5 was applied to the hidden layers.19Srivastava N. Hinton G. Krizhevsky A. Sutskever I. Salakhutdinov R. Dropout: a simple way to prevent neural networks from overfitting.J Mach Learn Res. 2014; 15: 1929-1958Google Scholar The training of RNN used batch-normalized images and persisted for 100 epochs, with the optimized model possessing the minimum loss being saved and implemented (Supplemental Table S2 and Supplemental Figure S1). To address the unbalanced frequency of Silva A, B, and C, weights w0 and w1 were used for the negative and positive classes, respectively, to assign greater significance to the underrepresented examples. In the training process, false-negative errors were given greater weight to produce models with heightened sensitivity. Two pathologists [J. Li (10 years of experience) and J.Y. (15 years of experience)] independently and blindly diagnosed WSIs in a single session. Subsequently, their diagnostic results were collected for accuracy evaluations and comparisons with Silva3-AI. Before conducting a quantitative evaluation of the Silva pattern classification results, a visual assessment of the heatmaps generated by Silva3-AI was executed. The heatmaps were artificially colored and presented alongside the original H&E slides. A qualitative assessment of the heatmaps of the predicted Silva pattern and its top representative patches with the highest attention scores assigned by Silva3-AI was done for all patients by two expert gynecologic pathologists (Xi. Zhang and Chunyan Hao). The raw WSIs were examined by using NDP.view2 (provided by WISLEAP, the scanner vendor), Case-Viewer (provided by 3DHISTECH, the scanner vendor), and uPath (provided by Roche Diagnostics, the scanner vendor). Image extraction and analysis were performed by using Pyvips (version 1.12.3) and OpenCV (version 4.1.1)20Culjak I, Abram D, Pribanic T, Dzapo H, Cifrek M: A brief introduction to OpenCV. Proceedings of the 35th International Convention MIPRO, Opatija, Croatia, 2012. pp. 1725–1730.Google Scholar in Python (version 3.6.6). The primary computing platform was a high-performance node equipped with two NVIDIA Tesla A100 GPUs (eight cores, 128 GB memory per core), and the deep learning model was developed by using PyTorch version 1.2.0.21Gould S. Hartley R. Campbell D. Deep declarative networks.IEEE Trans Pattern Anal Mach Intell. 2022; 44: 3988-4004PubMed Google Scholar Scikit-learn (version 0.21.2) and Matplotlib (version 2.2.2) in Python undertook major estimation and visualization work cooperatively. Analyses were conducted by using STATA Release 14 (StataCorp., College Station, TX). Statistical analysis was performed by using SPSS version 25.0 (IBM, Armonk, NY). Categorical variables are presented as numbers and percentages, and between-group differences were assessed by using the χ2 test. Normality tests were performed for all continuous variables, and all variables exhibited a normal distribution. Therefore, all continuous data are presented as the median (standard deviation) and were assessed by using the t-test. P < 0.05 was considered indicative of a statistically significant difference. The initial data set comprised 202 H&E–stained tissue slides collected from 202 patients with EAC at QLHSDU, consisting of 32 Silva A pattern slides, 42 Silva B pattern slides, and 128 Silva C pattern slides. These slides were randomly divided into training (60%, n = 121 slides) and internal testing (40%, n = 81 slides) subsets. Subsequently, the WSIs were cropped into small patches of 1024 × 1024 pixels at varying magnifications of 20×, 50×, 100×, 200×, and 400×. The total number of patches generated was 2,348,251, of which 1,408,951 were included in the training set, and 939,300 were reserved for internal testing to assess the classification performance of Silva3-AI (Supplemental Table S3). The clinicopathologic characteristics are presented in Table 2.Table 2Descriptive Clinicopathologic Characteristics of Patients with EACs in QLHSDU Cohort (N = 202) and Multicenter Independent Cohort (N = 161)CharacteristicQLHSDU cohort (N = 202)Multicenter independent cohort (N = 161)P valueTraining set (n = 121)Internal testing set (n = 81)Age, mean ± SD, y47.29 ± 10.0247.02 ± 9.3147.32 ± 9.780.342Tumor size, mean ± SD, cm3.08 ± 1.542.87 ± 1.302.82 ± 1.270.421Depth of invasion, n (%)0.275 <1/268 (56.43)43 (53.09)82 (50.93) ≥1/253 (43.57)38 (46.91)79 (49.07)FIGO stage, n (%)0.157 IA26 (4.96)4 (4.94)10 (5.98) IB140 (33.06)27 (32.92)45 (28.54) IB236 (29.75)24 (29.63)49 (30.5) IB320 (16.53)14 (16.56)24 (14.76) IIA19 (7.44)6 (7.51)19 (11.77) IIA210 (8.26)6 (8.44)14 (8.45)Differentiation, n (%)0.069 High differentiation23 (19.42)17 (20.99)38 (23.54) Middle differentiation50 (41.18)30 (37.04)71 (43.98) Low differentiation48 (39.4)34 (41.97)52 (32.48)Silva pattern, n (%)0.631 Pattern A19 (15.66)13 (1604)33 (20.61) Pattern B25 (20.87)17 (20.99)39 (24.35) Pattern C77 (63.47)51 (62.97)89 (55.04)LVI, n (%)0.467 Yes32 (26.42)21 (25.93)51 (31.74) No89 (73.58)60 (74.07)110 (68.26)LNM, n (%)0.742 Positive28 (19.18)19 (23.46)41 (25.36) Negative93 (77.02)62 (76.54)120 (74.64)HPV test method, n (%)0.086 HC 236 (29.75)25 (30.86)44 (27.33) PCR85 (70.25)56 (69.14)110 (68.32)Cytopathology features or p16007 (4.35)EAC, endocervical adenocarcinoma; FIGO, International Federation of Gynecology and Obstetrics; HC 2, Hybrid Capture 2; HPV, human papillomavirus; LNM, lymph node metastasis; LVI, lymphovascular invasion; PCR, polymerase chain reaction. Open table in a new tab EAC, endocervical adenocarcinoma; FIGO, International Federation of Gynecology and Obstetrics; HC 2, Hybrid Capture 2; HPV, human papillomavirus; LNM, lymph node metastasis; LVI, lymphovascular invasion; PCR, polymerase chain reaction. The performance of the model trained at the same magnification in the training set was evaluated on the internal test set. The results indicate that Silva C achieved the highest accuracy with an area under the receiver-operating characteristic curve (AUROC) of 0.964 when the magnification was set at 50× and the Top-K was 1000 number of patches. Similarly, when the magnification was set at 200× and the Top-K was 1000 number of patches, Silva B achieved the highest accuracy with an AUROC of 0.939, and Silva A achieved the highest accuracy with an AUROC of 0.923, as presented in Table 3, Table 4, Figure 5, and Supplemental Table S4.Table 3AUC of the QLHSDU Internal Test Set with Multi-MagnificationPatternMagnification∗Eyepiece was 10×; multi-magnification of objective was 2×, 5×, 10×, 20×, and 40×, respectively.20×50×100×200×400×Silva A0.5130.5560.6490.9230.871Silva B0.5160.5380.7530.9390.837Silva C0.7540.9640.8380.7120.694Bold indicates the highest performance.AUC, area under the curve; QLHSDU, Qilu Hospital of Shandong University.∗ Eyepiece was 10×; multi-magnification of objective was 2×, 5×, 10×, 20×, and 40×, respectively. Open table in a new tab Table 4Macro_F1_score and BA_score of the QLHSDU Internal Test Set with Variances Top-K Number of PatchesMetricTop-K (no. of patches)1050100500100015002000Macro_F1_score0.6930.7140.7240.8830.9170.8670.749BA_score0.6730.6950.7670.8710.9160.8340.714Bold indicates the highest performance.QLHSDU, Qilu Hospital of Shandong University. Open table in a new tab Bold indicates the highest performance. AUC, area under the curve; QLHSDU, Qilu Hospital of Shandong University. Bold indicates the highest performance. QLHSDU, Qilu Hospital of Shandong University. The predictive capability of Silva3-AI was significant in the internal test set. To process the multicenter independent cohort test set, the same structure and procedure as the internal testing set were used. This approach resulted in AUROC of 0.947, 0.908, and 0.947 for Silva C, B, and A, respectively. The 3 × 3 confusion matrices for the independent testing cohort provided a visual representation of the misclassifications made by Silva3-AI (Figure 6 and Supplemental Table S5). The heatmaps provided a comprehensive overview of slide predictions, and were used to identify underlying histopathologic patterns and simplify result interpretations. As shown in Figure 7, the fourth column consisted of the patches in the red boxes, which are the representative positive patches (high attention) of correctly classified slides, displaying the highest-relevant histopathologic characters of Silva A, B, and C patterns, respectively. In patches of Silva A, representative positive patches displayed proliferation of ne