Pd36-05 Deep Learning Predictive Model for Prostate Cancer Recurrence Using Clinical and Histological Data From the Tcga Database

You have accessJournal of UrologySurgical Technology & Simulation: Artificial Intelligence III (PD36)1 May 2024PD36-05 DEEP LEARNING PREDICTIVE MODEL FOR PROSTATE CANCER RECURRENCE USING CLINICAL AND HISTOLOGICAL DATA FROM THE TCGA DATABASE Petronio Augusto de Souza Melo, Gabriel Arantes, William Carlos Nahas, and Katia Ramos Moreira Leite Petronio Augusto de Souza MeloPetronio Augusto de Souza Melo , Gabriel ArantesGabriel Arantes , William Carlos NahasWilliam Carlos Nahas , and Katia Ramos Moreira LeiteKatia Ramos Moreira Leite View All Author Informationhttps://doi.org/10.1097/01.JU.0001008916.72488.6a.05AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookLinked InTwitterEmail Abstract INTRODUCTION AND OBJECTIVE: Predicting tumor behavior is a primary goal in oncology to guide patient management. Prostate cancer (PCa) is the leading cause of cancer in men in western countries. Adjuvant treatment in high-risk PCa after radical prostatectomy (RP) is related to important side effects, and there are few predictive markers to help urologists for its indication. The commercially available tests are based on molecular panels, that are expensive and not widely available. The use of artificial intelligence has been applied in diagnosis and grading of PCa, but are rarely used to predict the risk for tumor recurrence. This study aimed to combine clinical data with histological images from RP specimens to develop a deep learning model that predicts prostate cancer recurrence, using the data and the histological images of the The Cancer Genome Atlas (TCGA). METHODS: We employed data from 495 PCa patients who underwent RP from the public TCGA database: 403 were cured, and 92 experienced biochemical recurrence. Maintaining this cured-to-recurrence ratio, we randomly choose 396 patients for training and 99 for testing. Clinical parameters included age, tumor laterality, primary Gleason pattern, total Gleason score, tumor stage, and margin status. Histopathological slides, presented in svs format, were segmented into 2056 x 2056 pixel tiles. We used the EfficientNetB7 architecture for feature extraction. Following this extraction, a logistic regression classifier was employed, with features standardized prior to model training. The classifier was trained using a large number of iterations to ensure convergence. RESULTS: In the validation set, the model reported an accuracy of 82.96%. When evaluated on the test set, it delivered an accuracy of 81.74%. Specifically, the model exhibited a sensitivity of 24.8% in correctly identifying recurrences and a specificity of 95.6% for identifying cures. The precision for identifying recurrences was 57.8%, with a false positive rate of 4.4%. The ROC curve showcased an area of 0.74, attesting to the model's predictive capability. CONCLUSIONS: Our deep learning model, utilizing both clinical and histological data, demonstrated promising accuracy in predicting PCa recurrence post-surgery. The inclusion of important data, like PSA levels, not available in TCGA data set, could add more accuracy to the model. With this preliminary data we believe that this integrated approach may offer valuable insights for post-operative patient management and risk stratification, helping the difficult decision to introduce adjuvant therapy. Source of Funding: None © 2024 by American Urological Association Education and Research, Inc.FiguresReferencesRelatedDetails Volume 211Issue 5SMay 2024Page: e794 Advertisement Copyright & Permissions© 2024 by American Urological Association Education and Research, Inc.Metrics Author Information Petronio Augusto de Souza Melo More articles by this author Gabriel Arantes More articles by this author William Carlos Nahas More articles by this author Katia Ramos Moreira Leite More articles by this author Expand All Advertisement PDF downloadLoading ...