High-grade extrauterine serous carcinoma (HGSC) is an aggressive tumor with high rates of recurrence, frequent chemotherapy resistance, and overall 5-year survival of less than 50%. Beyond determining and confirming the diagnosis itself, pathologist review of histologic slides provides no prognostic or predictive information, which is in sharp contrast to almost all other carcinoma types. Deep-learning based image analysis has recently been able to predict outcome and/or identify morphology-based representations of underlying molecular alterations in other tumor types, such as colorectal carcinoma, lung carcinoma, breast carcinoma, and melanoma. Using a carefully stratified HGSC patient cohort consisting of women (n = 30) with similar presentations who experienced very different treatment responses (platinum free intervals of either ≤ 6 months or ≥ 18 months), we used whole slide images (WSI, n = 205) to train a convolutional neural network. The neural network was trained, in three steps, to identify morphologic regions (digital biomarkers) that are highly associating with one or the other treatment response group. We tested the classifier using a separate 22 slide test set, and 18/22 slides were correctly classified. We show that a neural network based approach can discriminate extremes in patient response to primary platinum-based chemotherapy with high sensitivity (73%) and specificity (91%). These proof-of-concept results are novel, because for the first time, prospective prognostic information is identified specifically within HGSC tumor morphology.

An artificial intelligence (AI) algorithm for prostate cancer detection and grading was developed for clinical diagnostics on biopsies. The study cohort included 4221 scanned slides from 872 biopsy sessions at the HUS Helsinki University Hospital during 2016–2017 and a subcohort of 126 patients treated by robot-assisted radical prostatectomy (RALP) during 2016–2019. In the validation cohort (n = 391), the model detected cancer with a sensitivity of 98% and specificity of 98% (weighted kappa 0.96 compared with the pathologist’s diagnosis). Algorithm-based detection of the grade area recapitulated the pathologist’s grade group. The area of AI-detected cancer was associated with extra-prostatic extension (G5 OR: 48.52; 95% CI 1.11–8.33), seminal vesicle invasion (cribriform G4 OR: 2.46; 95% CI 0.15–1.7; G5 OR: 5.58; 95% CI 0.45–3.42), and lymph node involvement (cribriform G4 OR: 2.66; 95% CI 0.2–1.8; G5 OR: 4.09; 95% CI 0.22–3). Algorithm-detected grade group 3–5 prostate cancer depicted increased risk for biochemical recurrence compared with grade groups 1–2 (HR: 5.91; 95% CI 1.96–17.83). This study showed that a deep learning model not only can find and grade prostate cancer on biopsies comparably with pathologists but also can predict adverse staging and probability for recurrence after surgical treatment.

Lung cancer is the leading cause of cancer-related deaths worldwide, and genetically-engineered mouse models (GEMMs) of cancer provide important mechanistic and preclinical insights into this deadly disease. In particular, the “KP” model enables lung-specific inducible activation of oncogenic Kras G12D, and loss of Trp53, the two most common driver events of human non-small cell lung cancer (NSCLC). Importantly, the KP model is widely used and faithfully recapitulates molecular and histopathological features of the human disease, including progression from early hyperplasia and adenoma to invasive adenocarcinoma. However, the KP model results in multi-focal and heterogeneous tumor burden, and there is a need for improved tools to increase throughput and decrease subjectivity of tumor burden quantification and histopathological analyses. To this end, we trained a convolutional neural network (CNN) for semantic multi-class segmentation using the Aiforia(R) platform. The CNN was trained to classify and detect lung parenchyma, NSCLC tumors, and NSCLC tumor grades (grade 1-4). For supervised training, we used selected areas from 93 hematoxylin and eosin stained slides. For validation, we analyzed 34 slides completely independent of the CNN training. Tumor grades were manually annotated on the validation slides blinded to the CNN results. The overall F1 score of the CNN in grade classification was 98% and total area error 0.3%. The grade-specific F1-scores were 89%, 97%, 99%, and 98% for grades 1, 2, 3, and 4, respectively. Corresponding grade-specific total area errors were 0.4%, 0.2%, 0.4%, and 0.1%. Manual scoring independent of the training and CNN yielded similar tumor burden and grading results. In addition, the algorithm accurately recapitulates the increased tumor burden and grade seen in KP tumors harboring additional mutation of the tumor suppressor Keap1, and the delayed kinetics of KP tumors harboring a strong T cell antigen, in independent datasets. We have also extended this methodology to identification of tumors in a GEMM of small cell lung cancer, a distinct class of lung cancer with poor prognosis. In conclusion, we demonstrate that deep neural networks can be used for automated analysis and grading of preclinical models of lung cancer. We anticipate that this powerful technology will increase the throughput, sensitivity and reproducibility of hypothesis-driven studies of factors influencing tumor progression and immune response in mouse models of lung cancer. Citation Format: Peter Maxwell Kienitz Westcott, Tuomas Pitkänen, Sami Blom, Thomas Westerling, Tuomas Ropponen, Nathan Sacks, Katherine Wu, Roderick Bronson, Tuomas Tammela, Tyler Jacks. Deep neural network for automatic histopathologic analysis of murine lung tumors [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2019; 2019 Mar 29-Apr 3; Atlanta, GA. Philadelphia (PA): AACR; Cancer Res 2019;79(13 Suppl):Abstract nr 4447.

No abstract