Editors: Liron Pantanowitz, MD, PhD, MHA, chair of the Department of Pathology and professor of pathology, University of Pittsburgh Medical Center, and Matthew G. Hanna, MD, vice chair of pathology informatics and associate professor, Department of Pathology, University of Pittsburgh Medical Center.
Deep learning-based analysis of gross features of OETs
July 2025—Computational pathology has largely focused on analyzing tissue slides and overlooked worthwhile information in gross images. Recognizing this, the authors introduced a novel deep learning model based on Swin Transformer architecture and called Swin Transformer-based Gross Features Detective Network (SGFD-network). They sought to summarize and classify the gross examination characteristics of ovarian epithelial tumors by analyzing images of gross specimens using this deep learning approach. The SGFD-network is especially useful for distinguishing borderline tumors with microinvasive components from frank carcinomas. This capability is crucial during frozen section analysis, in which limited sampling and time pressures often challenge diagnostic accuracy. The SGFD-network was trained on a data set of 4,129 curated gross images covering five common histologic subtypes of ovarian epithelial tumors: serous, mucinous, seromucinous, endometrioid, and clear cell carcinoma. It demonstrated superior performance in distinguishing borderline tumors from carcinomas when compared to radiological and ultrasound diagnosis, achieving a classification accuracy of 88.9, 86.4, and 93.0 percent for the benign, borderline, and carcinoma groups, respectively. For serous tumors, the SGFD-network had accuracy rates of 96.1, 76.1, and 95.8 percent for benign, borderline, and carcinoma cases, respectively. And for mucinous tumors, it had accuracy rates of 90.1, 87.5, and 67.9 percent for benign, borderline, and carcinoma cases, respectively. (Mucinous tumors are often difficult to assess grossly due to their heterogeneous morphology.) Notably, the model could detect microinvasive foci, which are typically challenging to distinguish by gross examination alone. The model demonstrated superior performance compared with frozen section, radiological, and ultrasound diagnosis in detecting cases with microinvasion, achieving a 92.2 percent accuracy rate. It showed heat map visualization and unsupervised clustering of gross features, including cystic or solid content, surface texture, and color. These visualization tools provide insight into the model’s decision-making and enhance its potential as a real-time aid in the frozen section room. The SGFD-network is also computationally efficient, requiring minimal graphics processing unit resources and yielding predictions in less than one second per image. The authors concluded that by continuing to develop and refine the SGFD-network, it could become an effective tool for improving gross examination and intraoperative sampling of ovarian epithelial tumors.
He D, Jin L, Geng H, et al. Deep learning-based analysis of gross features for ovarian epithelial tumors classification: A tool to assist pathologists for frozen section sampling. Hum Pathol. 2025. doi.org/10.1016/j.humpath.2025.105762
Correspondence: Dr. Lanqing Cao at caolanqing@jlu.edu.cn
PIRO: a Web-based search platform for pathology reports
A complaint of pathologists is the difficulty of obtaining archival data in the laboratory information system. This is problematic because pathologists typically use such data for clinical, quality assurance, research, and educational purposes. Much of this laboratory-related data is unstructured and contained solely within text. This presents issues because many laboratory information systems (LISs) have only rudimentary functionality for conducting text searches. To address this issue, researchers at Cleveland Clinic developed a Web-based platform called PIRO (Pathology Information Retrieval Optimizer). Using PIRO, nontechnical staff in the pathology lab and other areas of Cleveland Clinic efficiently searched text-based, pathology-generated reports in the institution’s Epic Beaker LIS and its previous LIS, Cerner CoPath. (Legacy pathology reports in CoPath had been extracted, transformed into PIRO schema, and loaded into Cleveland Clinic’s SQL server database.) The PIRO SQL database holds millions of surgical pathology, cytopathology, bone marrow biopsy, and autopsy reports. PIRO offers such capabilities as faceted filtering, data extraction, case cohort building, and search result export. Faceted filtering allows users to narrow down search results by applying multiple filters, and users can save search results in Microsoft Excel files. PIRO provides secure access control within the institutional network, so it complies with institutional privacy protocols. The developers of PIRO also deployed a large language model to annotate reports with a “malignancy risk” label, which further enhances the precision of text searches. In an eight-month study of PIRO conducted at Cleveland Clinic in 2024, the platform averaged 12,288 queries per month. Eighty of 119 (67 percent) AP staff pathologists performed at least one PIRO search during that time, and only 11 (nine percent) used the Epic Beaker SlicerDicer self-service data-exploration tool instead. Given the growing demand for data mining and analytics in pathology, the authors concluded that other institutions may want to consider using PIRO for advanced search and retrieval of information in their pathology archives.
Robertson S, Koppireddy V, Cumbo J, et al. PIRO: A web-based search platform for pathology reports, leveraging large language models to generate discrete searchable insights. J Pathol Inform. 2025. doi.org/10.1016/j.pi.2025.100436
Correspondence: Dr. Scott Robertson at roberts10@ccf.org