Weakly supervised deep learning-based detection of serous tubal intraepithelial carcinoma in fallopian tubes

Abstract

Serous tubal intraepithelial carcinoma (STIC) is an uncommon, non-invasive carcinoma that occurs more frequently in individuals with germline BRCA mutations and is an established precursor to high-grade serous ovarian carcinoma. STIC can be challenging to detect during pathologist evaluation, as it can manifest as a small focus of atypia in an otherwise benign salpingectomy specimen. There is a clinical need for scalable, weakly supervised computational approaches to aid in the detection of STIC. We developed a deep learning model to identify STIC and serous tubal intraepithelial lesions (STIL) in whole-slide images. We obtained fallopian tube specimens diagnosed as STIC (n = 49), STIL (n = 48), and benign fallopian tube (n = 83) at a single academic medical center. We trained a weakly supervised, attention-based multiple instance learning model and evaluated performance on independent datasets, including an additional unbalanced dataset (n = 40 benign, n = 2 STIL, n = 1 STIC) and cases diagnosed descriptively as benign reactive atypia (n = 53). The model achieved high sensitivity and specificity on the balanced validation cohort, with an area under the receiver operating characteristic curve (AUROC) of 0.96 (95% CI: 0.90–1.00), and demonstrated similarly strong performance on unbalanced validation cohorts (AUROC 0.98). Interpretability analyses indicated that model decisions were based on epithelial atypia. These results support the potential of integrating deep learning screening tools into clinical workflows to augment pathologist efficiency and diagnostic accuracy in fallopian tubes.

Introduction

High-grade serous carcinoma of the ovary (HGSC) is the most common subtype of epithelial ovarian cancer and is the fifth leading cause of cancer-related deaths in women.¹ Approximately 15–25% of HGSC cases are associated with germline mutations in BRCA1/BRCA2, and nearly all of HGSC exhibit somatic TP53 mutations.^2,3 HGSC is thought to originate in the fallopian tubes, arising from serous tubal intraepithelial carcinoma (STIC), a microscopic, non-invasive carcinoma of the fallopian tube epithelial lining, based on shared TP53 mutations between STIC and concurrent HGSC.^4,5 Whereas STIC is rare in the general population, occurring in fewer than 1% of salpingectomies,⁶ it is more frequently identified in BRCA mutation carriers, with reported rates of 3–5% of risk-reducing salpingo-oophorectomies.⁷ As such, incidental detection of STIC may prompt consideration of BRCA mutation testing in patients without known genetic risk factors.⁸

Diagnosis of STIC is based on a combination of histopathological and immunohistochemical criteria.⁹ Morphologically, STIC is characterized by nuclear enlargement, loss of polarity, hyperchromasia, nucleolar prominence, mitotic activity, epithelial tufting, and absence of cilia. However, some features can overlap with benign or reactive tubal epithelium; therefore, immunohistochemical staining for p53 and Ki-67 is now considered an essential part of the diagnosis. In the setting of morphological atypia, STIC is defined by aberrant expression of p53 and Ki-67 proliferation index ≥10% in the lesional cells. Cases with foci of morphological atypia and aberrant p53 expression but Ki-67 proliferation index <10% are diagnosed as serous tubal intraepithelial lesion (STIL), which currently has unclear clinical implications.¹⁰

Despite its clinical significance, diagnosing STIC can be difficult and is fraught with high interobserver variability.^8,9 In addition to their rarity, STIC lesions can be very small relative to the benign epithelium, which is composed of multiple cell types and can exhibit various degrees of baseline atypia. Atypia itself is a relatively subjective finding, reflected in studies that have found Cohen's kappa values of 0.67 and 0.27 for STIC and STIL, respectively.⁹ Although consensus criteria and use of immunostains have improved diagnostic consistency, some variability persists.⁹ Moreover, protocols such as SEE-FIM (Sectioning and Extensively Examining the Fimbria) that are designed to enhance lesion detection increase the histological workload per case.¹¹

These challenges underscore the need for scalable computational approaches to aid in the detection of STIC. However, few studies have employed machine learning models for this task. A recent study by Bogaerts et al. developed a highly accurate model for detection of STIC/STIL versus benign epithelium using a large multi-institutional dataset and extensive manual annotation of regions of interest (ROIs) during model training.¹² Evaluation with model assistance improved pathologist sensitivity and speed.¹³ However, strongly supervised model training approaches may limit scalability with additional data or adapting a model to institution-specific datasets.¹⁴

In this study, we developed a weakly supervised deep learning model for classifying fallopian tube specimens as benign or STIC/STIL using whole-slide images (WSIs). Our approach does not rely on manual annotation of ROIs, facilitating expansion to larger training or clinical datasets. These findings support the integration of deep learning-based screening tools into diagnostic workflows to improve the detection and classification of early serous neoplasia in the fallopian tube.

Results

We collected cases of STIC and STIL from the University of Michigan Department of Pathology archives from 2013 to 2024 and scanned H&E WSIs at 40× magnification (0.25 μm/pixel). We randomly split the cohort into a training dataset (70%, n = 126) and evaluation dataset (30%, n = 54), each balanced for positive and negative cases (Fig. 1). The median patient age at diagnosis of STIC and STIL cases was 62 years (IQR 55–69) compared with 34 years (IQR 30–38) in the benign cases. There were relatively few cases with positive mutation status for BRCA1/BRCA2 or other risk-increasing mutations (n = 11), most of which were randomly assigned to the training group, precluding subgroup analyses by mutation status. From each WSI, we extracted tiles from areas containing epithelium and applied multiple normalization methods to reduce background and adjust for differences in staining (see Methods). We trained an attention-based multiple instance learning (aMIL) classifier built on the pathology-specific foundation encoder UNI with 3-fold cross-validation.

Fig. 1.

Diagram of study workflow. (A) H&E slides were scanned at 40× magnification and randomized to training and hold-out evaluation groups. Epithelium was segmented using a U-Net model. Tiles were extracted with normalization methods. Tiles were converted to feature vectors with the UNI foundation model and aggregated into bags for each slide. An attention-based multiple instance learning model was trained based on these bags of feature vectors and evaluated on the held-out evaluation dataset. (B) WSI thumbnail of an example case. H&E slides were scanned at 40× magnification. (C) Areas containing epithelium were segmented and saved as ROI. (D) Tiles were extracted from these ROI and used in downstream model training.

We assessed model performance on the held-out evaluation dataset, which was not seen by the model during training. The model exhibited strong performance based on several metrics. Area under the receiver operating characteristic curve (AUROC) was 0.96 (95% CI: 0.90–1.00; Fig. 2A). At a threshold of 0.3, the sensitivity was 0.90 with a specificity of 0.96, corresponding to precision of 0.96, accuracy of 0.93, and F1-score of 0.93. At a threshold of 0.5, the sensitivity was 0.83 with a specificity of 1.00, corresponding to precision of 1.00, accuracy of 0.91, and F1-score of 0.93. Based on a disease prevalence of 1%, these test characteristics correspond to a positive-predictive value (PPV) of 19% and negative-predictive value (NPV) of 99.9%. With a disease prevalence of 5%, the model PPV is 54% with a NPV of 99.5%. To more closely examine model predictions, we plotted the prediction score by diagnosis category (Fig. 2B). Prediction scores for STIC and STIL were significantly higher than those for benign cases (median 0.99 vs. 0.04, p value <0.001). There was no significant difference in prediction score between STIC and STIL cases (median 0.97 vs. 0.99, p value 0.16).

Fig. 2.

Model performance and interpretability. (A) Model performance shown as a ROC curve (orange). Bootstrap ROC curves are shown (dark blue). (B) Scatterplot of prediction scores on the evaluation cohort. Each case is shown as a point, with the diagnosis group on the x-axis and the prediction score on the y-axis. Boxplots display the median and interquartile range with the whiskers showing the farthest point up to 1.5 times the interquartile range. The dotted line denotes a threshold score of 0.3. (C, D) Examples of a false-negative cases, including H&E (left), p53 IHC, and Ki-67 IHC. The images are shown at 15× magnification. Panel C shows a case from a known BRCA mutation carrier with a prediction score of 0.03. Panel D shows an incidental STIC diagnosis with prediction score of 0.07. (E) UMAP embedding of tile feature vectors shown as a density contour plot, with tiles originating from benign cases shown in light blue, STIL cases shown in dark blue, and STIC cases shown in orange. (F) UMAP embedding plotted as H&E tile mosaic. (G) Inset of tile mosaic in benign area. (H) Inset of tile mosaic in STIC/STIL area. (I) Attention scores shown as heatmap over WSI. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

There were few outliers in each diagnosis group, including three false negatives in the STIC group using a threshold score of 0.3. Examination of these cases revealed that each lesion was described in the original diagnostic report as a small or minute focus. Examination of the relevant foci showed qualitatively small lesions (Fig. 2C, D). Two of the false-negative cases were associated with known BRCA mutation. The third false negative was a case of incidental STIC; the patient was referred for genetic counseling and underwent bilateral oophorectomy with benign pathological findings.

We assessed model characteristics with dimension–reduction methods and visualization of attention scores. Uniform manifold approximation and projection (UMAP) embedding of the final model activation layer showed subclusters of STIC and STIL features separate from those of the benign cases (Fig. 2E). Examination of a UMAP tile mosaic showed tiles with increased atypia in the STIC/STIL regions relative to the benign regions (Fig. 2F–H). Examination of attention heatmaps showed higher attention scores over epithelial surfaces compared to stroma (Fig. 2I). Together, these results suggest that the model decisions are based on assessment of relevant histological features.

Because STIC and STIL are relatively rare diagnoses, we sought to evaluate the model's performance on an unbalanced validation dataset. We collected an additional cohort of cases consisting of 40 benign cases and 3 STIC/STIL cases. The model demonstrated strong performance on this cohort with an AUROC of 0.98 (95% CI: 0.90–1.00, Fig. 3A). At a threshold of 0.3, there was one false positive and no false negatives (Fig. 3B). These results suggest that the model was not significantly overfit on the training dataset.

Fig. 3.

Validation on additional cohorts. (A) Model performance on the unbalanced cohort shown as an ROC curve. (B) Scatterplot of prediction scores on the unbalanced cohort. Each case is shown as a point, with the diagnosis group on the x-axis and the prediction score on the y-axis. Boxplots display the median and interquartile range with the whiskers showing the farthest point up to 1.5 times the interquartile range. The dotted line denotes a threshold score of 0.3. (C) Scatterplot and boxplot of prediction scores on benign reactive atypia cases. The dotted line denotes a threshold score of 0.3.

We applied the model to a cohort of cases diagnosed descriptively as benign reactive atypia (n = 53). This is a heterogeneous set of cases that the interpreting pathologist subjectively identified as containing some degree of morphological atypia, but for which staining patterns for p53 and Ki-67 did not meet criteria for STIC or STIL and therefore did not fit into a single discrete biological entity. Consistent with this expected heterogeneity, analysis of the prediction scores showed a wide range of prediction scores (median 0.77, SD 0.40), with 34% of the cases below a prediction score threshold of 0.3 (Fig. 3C). These results suggest that although the current model may be partially able to distinguish reactive from neoplastic atypia, more exuberant atypia requires immunohistochemical staining to resolve.

Discussion

In this study, we developed and evaluated a deep learning model to classify fallopian tube WSI as benign or containing STIC/STIL. Despite a relatively modest sample size drawn from a single institution, the model achieved high performance metrics based on H&E alone and demonstrated robust generalization across independent test sets. Whereas understanding model behavior is a persistent challenge in computational pathology, interpretability analyses suggested that this model focuses attention on the epithelium, particularly on areas of nuclear atypia. These findings highlight the potential of deep learning tools to assist pathologists in screening workflows for cancer risk lesions.

A key strength of this study is the weakly supervised approach without the use of pathologist ROI annotations within WSI. ROI annotation is a labor-intensive process and itself can be a subjective process with interobserver variability. Whereas ROI annotation can help direct model attention to relevant epithelial areas, it complicates the process of updating models with additional data. By relying solely on slide-level labels, our model architecture enables more scalable and flexible dataset expansion as new cases are diagnosed.

This study has several limitations. The cases were collected from a single institution, and the sample size is relatively small due to the rarity of these lesions. It is unclear to what extent the model predictions would generalize to other institutions. To mitigate these concerns, we implemented rigorous cross-validation, stain normalization, and the use of pathology-specific pre-trained encoders. Despite the modest sample size, the model exhibited strong performance on multiple cohorts not seen in the training process, achieving similar AUROC as studies with larger sample size focused on STIC/STIL detection.¹² It is important to note that our weakly supervised approach offers benefits of scalability, but not necessarily stronger predictive power compared to other model architectures.

A future challenge for the field is to expand to multi-class identification. Although the model presented by Bogaerts et al. and the model presented here are unable to distinguish between STIC and STIL, future models with additional data may be able to more definitively distinguish STIC, STIL, and benign findings such as reactive atypia. A model that can distinguish STIC versus STIL would be particularly useful because only STIC has definitive clinical implications. Development of models to distinguish STIC/STIL versus benign reactive atypia will likely be challenging, as it is difficult or impossible even for subspecialized gynecological pathologists without the use of immunohistochemical stains. We expect institutional sample sizes to increase as digital pathology is more widely adopted, which may increase model power for multi-class identification.

Deep learning models for screening and detection of STIC face several challenges with clinical adoption. All models face a trade-off between sensitivity and specificity. In practice, a model with a high-sensitivity threshold could act as a screening or triage tool, flagging potentially positive cases for expedited pathologist review. If sufficiently high NPV are achieved, such models may help reduce time spent reviewing negative slides and augment pathologist efficiency in high-volume settings. However, even with a highly accurate classification model, the low incidence of STIC makes it challenging to achieve a high PPV. Computational models serve only as adjuncts to pathologists' clinical judgment, and individual pathologists may interact with model assistance in different ways.¹³ Deep learning models alone cannot solve issues of high interobserver variability in diagnosis, but with continued development, they may be able to increase pathologist efficiency and enhance patient safety by providing a second layer of detection. Even with the aid of computational models, diagnosing STIC remains a complex task requiring careful integration of the clinical context, histological findings, and immunohistochemical findings.

Methods

Research ethics and case identification

This research was approved by the University of Michigan Institutional Review Board. We obtained all diagnostic fallopian tube H&E slides available in the University of Michigan Department of Pathology archives from 2013 to 2024. We identified cases of STIC and STIL without concurrent neoplastic or inflammatory diagnoses within a given diagnostic slide. All cases were diagnosed by subspecialized gynecological pathologists at the University of Michigan; most cases were seen by at least two subspecialized gynecological pathologists for confirmation during the routine diagnostic process. All cases of STIC and STIL were confirmed by immunohistochemical staining for p53 (mouse monoclonal DO-7, Agilent) and Ki-67 (rabbit monoclonal 30-9, Roche-Ventana). We collected a set of benign fallopian tube specimens that were resected for non-neoplastic indications (e.g., desire for permanent sterilization). We stained a subset of these benign cases (n = 44) with p53 and found that all had wild type/non-aberrant p53 expression. Finally, we collected a set of fallopian tubes diagnosed as reactive atypia, which had foci of morphological atypia but lacked immunohistochemical aberrancies for p53 and Ki-67 (n = 53).

Model development

We obtained H&E WSIs at 40× magnification (0.25 μm/pixel) with a Leica-Aperio GT450 scanner. We randomly split the data into a training dataset (70%, n = 68 STIC/STIL, n = 58 benign) and a held-out evaluation dataset (30%, n = 29 STIC/STIL, n = 25 benign) not used in the model training process. Images were de-identified and analyzed on the University of Michigan Great Lakes High-Performance Computing cluster. The model training and evaluation workflow was performed using Slideflow version 3.0.2 with the cuCIM backend.¹⁵ We trained and applied a U-Net tissue segmentation model to identify areas of epithelium within each WSI. We extracted tiles from these segmented areas at 40× magnification and 512 × 512 pixels. For each tile, we performed grayspace filtering, Otsu's thresholding, and Gaussian blurring to remove background and reduce interference from artifacts such as pen marks. We applied stain normalization using the Reinhard algorithm.¹⁶ After tile extraction and normalization, we converted tiles into feature vectors using the UNI encoder, a domain-specific pre-trained foundation model for pathology.¹⁷ We trained aMIL models based on these feature bags for binary classification of benign or STIC/STIL with a PyTorch backend. We used the following model training parameters: learning rate of 1e-4, batch size n = 32, bag size of 256, and 30 epochs; model training included image augmentation such as image rotation and flipping. We used 3-fold cross-validation within the training dataset. After cross-validation, the model was trained on the full training dataset.

Model evaluation

The performance of the final model was assessed on the evaluation dataset. WSI were processed in the same manner as the training dataset, including epithelium segmentation, tile extraction, and normalization. We calculated performance metrics of sensitivity, specificity, AUROC, precision, accuracy, and F1-score. We calculated 95% confidence intervals by bootstrapping with 1000 iterations. We plotted the prediction scores by the underlying diagnosis and compared binary classification groups with a Student's t-test. We plotted attention score heatmaps to visualize areas used most by the model for classification. We generated UMAP embeddings of the final activation layer and plotted as 2D density contours and tile mosaics. We evaluated the model on additional evaluation cohorts, including an unbalanced cohort (n = 40 benign cases, n = 3 STIC/STIL cases) and a cohort of reactive atypia cases (n = 53). H&E and immunohistochemical stains were examined for post-hoc analysis of false-negative cases; immunohistochemical stains were not used in model training or validation.

Data and code availability

Code for the analysis is available on GitHub at https://github.com/andrewvalesano/STIC_ML. Slideflow is open-source and is available on GitHub at https://github.com/slideflow/slideflow. De-identified whole slide images may be made available upon reasonable request.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this article.