ECP 2023 Abstracts

S237 Virchows Archiv (2023) 483 (Suppl 1):S1–S391 13 Z plane) and without Z stacks. A deep learning based nuclei detection and classification model was built, followed by training of class-aware nearest neighbour graph to evaluate the distance, direction and proxim- ity between different nuclei for both WSIs of cases. Results: In several of the cases, the WSIs with Z stacks yielded more accurate distance measurement and directional vector analysis as it enables real 3-D modelling of nuclei across Z stacks. The Delta for distance measurement on best focus image vs z stack image ranges to an extent of 2.5 microns. Z stacking also increases the overall nuclei counts across layers in the range of 3-4 %, thus further improving the accurate tumour microenvironment analysis. Conclusion: Capturing Z stacks for WSI generation enables more accurate tumour microenvironment analysis, as it allows 3D directional vector and modelling for more accurate distance analysis among differ- ent nuclei types, in addition it also increases the overall nuclei counts distributed across Z -layers. Further studies are warranted to evaluate how the distance and direction of nuclei affects the treatment response for breast and other solid tumours. E-PS-08-020 A deep learning based tool for histological detection of malignancy In prostate core needle biopsies R. Jialdasani*, S. Saraf, A. Khandelwal, P. Chatterjee, V. Bharani, R. Kumar, D. Kaur, B. Gogoi, B. Singh, M.A. Osama, F. Bhatt, A. Narwal, W.P.K. Teng, S. Karakaya, K.V. Santosh *Qritive AI India Private Limited, India Background & objectives: Diagnosing prostate adenocarcinoma on histopathology is complex and highly subjective. We have developed a deep learning-based assistive AI tool to aid practising pathologists to identify and classify malignancies using a segmentation-based output. Diagnosing prostate adenocarcinoma on histopathology is complex and highly subjective. We have developed a deep learning-based assistive AI tool to aid practising pathologists to identify and classify malignan- cies using a segmentation-based output. Methods: Ten pathologists, from 8 institutions were given 150 whole slide images (WSIs) of prostate core needle biopsies obtained from mul- tiple institutes. Review was done in two phases – without (Phase1) and with AI assist (Phase2) marking benign and malignant areas. Ground truth (GT) was established by 2 senior pathologists. A semantic seg- mentation algorithm producing a pixel-based segmentation was used. Results: The WSIs were hosted over a cloud server and reported digi- tally. The AI agreed with GT in 148 / 150 cases, with 2 false positives. The agreement (=>8 doctors) with GT for benign cases was 34/41 and 41/41 in Phase 1 and 2, while for malignant cases, it was 105/109 and 107/109 respectively. The mean concordance for benign cases after AI assist increased by 17.07% (34 to 41) and 1.9% (105 to 107) for malignant cases. The concordance score of one of the pathologists for identifying benign cases improved by 32% using AI assist and achieved a perfect accuracy of 100%. Conclusion: Our study demonstrates our AI based prostate cancer detection module showed a sensitivity of 100% for malignancy detec- tion and a specificity of 95.12% in whole slide images of prostate core needle biopsies. Such AI based assistive tools have a potential to improve the concordance in independently practising pathologists while reporting complex cases like prostate needle biopsies, thus improving health equity. We intend to further refine the model to include ISUP grading and perform clinical validations with larger datasets. E-PS-08-021 Stratipath Breast: deep learning-based risk stratification of inter- mediate risk breast cancers S. Kang Lövgren*, P. Weitz, J. Ho, K. Ledesma Eriksson, B. Su, Y. Wang, S. Robertson, J. Hartman, M. Rantalainen *Stratipath, Sweden Background & objectives: In current clinical routine >50% of breast cancers are assigned an intermediate risk (NHG 2), with limited clini- cal value in treatment decisions. Stratipath Breast is the first CE-IVD marked solution for AI-based histopathology risk stratification into low- and high-risk groups. Methods: We evaluated the prognostic performance of the predicted risk classes for 901 NHG 2 primary breast cancer patients, 204 origi- nating from the TCGA BRCA study and 697 from a Swedish study. Prognostic performance was assessed by computing hazards ratios for recurrence with Cox proportional hazards models. Results: In the present study the point estimate for the marginal haz- ards ratio between predicted low- and high-risk groups was found to be 2.2. Adjusting for clinical covariates yielded a hazards ratio of 2.1 across all patients. Conclusion: Stratipath Breast enables risk-stratification of intermedi- ate risk breast tumours into low- and high-risk groups, while signifi- cantly reducing costs and turn-around times compared to molecular diagnostics. Its integration into routine clinical workflows therefore has the potential to benefit both patients and healthcare providers and to expand access to precision diagnostics. E-PS-08-022 Convolutional artificial neural network-based recognition of pan- creatic adenocarcinoma by Stimulated Raman Scattering scanning microscopy in unstained tissue sections É. Kocsmár*, K. Ócsai, Z. Mucsi, B. Barkóczi, K. Borka, A. Pesti, E. Kontsek, A. Kiss, B. Rózsa, G. Lotz *Semmelweis University, Department of Pathology, Forensic and Insurance Medicine, Hungary Background & objectives: Stimulated Raman Scattering (SRS) scan- ning microscopy allows the label-free spectral fingerprinting of lipids, proteins and water in cells and tissues, thus enabling morphological examination and analysis of chemical composition of unstained histo- logical specimens, moreover recognition of malignancy based on this. Methods: SRS imaging was performed on unstained sections of 49 pancreatic ductal adenocarcinomas at 2895 cm-1 and 2950 cm-1 wave- length channels. After hematoxylin-eosin staining, the cancer cell per- centage and normal tissue components of the imaging site was deter- mined. 2133 SRS images were used to train a residual convolutional artificial neural network (ResNet-101), whose cancer detection ability was tested in 2106 images. Results: The detection rate for individual non-tumorous tissue catego- ries (including pancreatic exocrine and endocrine tissue, nervous tis- sue, lymphatic tissue, adipose tissue, connective tissue, smooth muscle, inflammatory infiltrate, chronic pancreatitis and mixed non-tumorous elements) ranged from 57.7% to 100%, with the overall detection rate of non-tumorous tissue reaching 87.8%. For images with more than 30% adenocarcinoma, the cancer detection rate was 97.0%, for tumour contents between 15% and 30% it was 80.4%-91.9%, and for tumour percentages below this level it was 53.6-56.4%. Overall, the cancer detection sensitivity was 92.1%, specificity 87.8%, positive and nega- tive predictive value 91.6% and 88.4%, respectively, using convolu- tional artificial neural network. Conclusion: On deparaffinized unstained sections of formalin-fixed paraffin-embedded tissue blocks, SRS imaging combined with convo- lutional artificial neural network detection was able to identify the pres- ence of adenocarcinoma with excellent results even when trained on a relatively small dataset. As the intra- and inter-institutional variability of unstained sections is much lower than that of conventional hema- toxylin-eosin stained sections, their use for imaging with advanced microscopy techniques is promising.