ECP 2023 Abstracts

S242 Virchows Archiv (2023) 483 (Suppl 1):S1–S391 13 Conclusion: Although there was no statistically significant associa- tion in the parameters analysed, an association was observed between staging and the parameters, nuclear density and positive intensity average. Nuclear size and less chromatic nuclei might be related to invasiveness. This work sought to contribute to the knowledge and development of applications of new technologies of image analysis and artificial intel- ligence in Medicine, and specifically in PA and NMIBC. E-PS-08-040 ChatGPT in pathology applications: harnessing AI language mod- els for diagnostic decision support S. Spasic*, D. Gonzalez *Department of Pathology, Beo-lab, Medicover, Belgrade, Serbia Background & objectives: ChatGPT, a popular large-scale language model, has shown potential in healthcare applications. We aim to explore its utility as an AI-assisted decision support tool in surgical pathology. Methods: We adapted ChatGPT-4 to provide diagnostic suggestions based on histopathologic descriptions. Experiments involved listing diagnostic entities (e.g. follicular neoplasms), differentiating entities (e.g. basal cell carcinoma vs. trichoepithelioma), and generating dif- ferential diagnoses for specific histologic descriptions (e.g. follicular neoplasms with desmoplastic stroma). Results: ChatGPT demonstrated promise in generating accurate diag- nostic suggestions, though with limitations. It listed many but not all entities within a category, failed to mention CK20 for distinguishing basal cell carcinoma from trichoepithelioma, and did not provide all relevant diagnostic suggestions for the given histologic descriptions. Despite these limitations, ChatGPT offered valuable insights, indi- cating potential for enhancing diagnostic accuracy and efficiency in pathology. Conclusion: ChatGPT shows promise as an AI-assisted decision sup- port tool in pathology applications, but its limitations warrant caution. These include generating false concepts (hallucination effect), outdated training data, and inability to challenge existing biases. ChatGPT holds potential to enhance diagnostic outcomes and streamline workflow in pathology. Future research should focus on refining the model, address- ing limitations, and integrating ChatGPT into clinical practice. E-PS-08-041 Digital and mobile H&E imaging of bulk tissue in the operating room M. Strauch*, J.P. Kolb, C. Rose, N. Merg, J. Hundt, C. Kümpers, S. Perner, S. Karpf, R. Huber *Medical Laser Center Lübeck, Germany Background & objectives: During surgery, frozen sections ensure the quality of the process. However, frozen section analysis can take up to 60 minutes depending on the availability of a pathologist, sample prep- aration time and transport delays. Meanwhile the surgery is interrupted. Methods: We develop a mobile multiphoton microscope prototype for the operating room (OR) that uses advanced laser technology to cap- ture H&E images without the need to freeze or section the bulk tissue sample. The tissue surface is quickly stained with H&E before being placed in the microscope. It is automatically scanned and the result is displayed in the OR or remotely. Results: We designed the microscope to meet the operating room’s requirements. We developed an H&E staining protocol that allows sample preparation and measurement in less than 15 minutes. We first investigated porcine samples and have now moved on to leftover tissue from plastic surgery to prove the concept. Measuring tissue samples directly in the OR saves transport time and therefore reduces the waiting time for the surgeon. Once the measure- ment is finished, the result can be displayed on the device’s monitor and on the pathologist’s PC. This improves the interaction between the surgeon and pathologist; both can view the same images zoom in and out and discuss the findings together. Conclusion: We demonstrated a microscopy technique, that is capa- ble of obtaining digital H&E images from bulk H&E-stained tissue without the need for freezing, sectioning, or glass slide preparation. We have developed a protocol for the workflow and improved our setup to a mobile platform to allow it to be shared between differ- ent operating rooms. The images can be examined remotely by the pathologist, eliminating transport time and thus speeding up the surgery. Funding: The project Multiphoton microscopy for section-free H&E histology is funded by the Federal Ministry for Economic Affairs and Climate Action and the European Social Fund as part of the EXIST program. E-PS-08-042 Developing an integrative model using histopathological images and clinico-genomic data for predicting the prognosis of patients with papillary renal cell carcinoma M.J. Tan*, S. Kee, M.A. Sy, S. Border, N. Lucarelli, A. Gupta, M. Masalunga, F.I. Ting, P. Sarder *University of Florida, USA Background & objectives: Overlapping histopathologic features in renal epithelial tumours require improving the criteria for evaluating papillary renal cell carcinoma (pRCC). We propose an integrative model (IM) that combines morphologic and clinico-genomic features to obtain better prognostication for pRCC. Methods: Matched histopathological images, genomic, and clinical data (race, AJCC tumour stage, and sex) from The Cancer Genome Atlas were used. Image feature extraction was done using CellProfiler. Prognostic image features were selected using least absolute shrink- age and selection operator, and support vector machine algorithms. Weighted gene co-expression network analysis was used to determine eigengene modules. Results: Risk groups based on prognostic features were significantly distinct (p < 0.05) according to Kaplan-Meier analysis and log-rank test results. Two image features and nine modules were used in Ran- dom Survival Forest models, measuring 11-, 16-, and 20-month areas under the curve (AUC) of a time-dependent receiver operating curve. The IM (AUCs: 0.86, 0.85, 0.87) outperformed models trained on eigengenes alone (0.75, 0.733, 0.785), morphological features alone (0.593, 0.523, 0.603), and clinical features alone (0.743, 0.757, 0.743). It also significantly outperformed the IM without clinical data (0.877, 0.769, and 0.811) in 16- and 20-month predictions. Conclusion: Image granularity and Zernike shape features, along with nine eigengene modules and three clinical features, were iden- tified as prognostic image and clinico-genomic features for patients with pRCC. Results suggest that an integrative model combining his- topathological images and clinico-genomic features could improve survival prediction, especially in longer timeframes. Enrichment anal- ysis also revealed associations with the NRF2 pathway and metabolic processes, corroborating the metabolic nature of pRCC. Therefore, the IM model shows potential applicability in clinical decision-mak- ing, particularly in personalized treatment regimens.