Multimodal deep learning models to predict endometrial cancer recurrence risk by integrating histopathology and genomic data

Andrew Polio; Samantha Metzger; Katherine Boecking; Vincent Wagner; David Bender; Michael Goodheart; Jesus Gonzalez Bosquet

doi:10.1016/j.ygyno.2025.04.434

Objectives Approximately 15–20 % of all patients with endometrioid endometrial cancer (EEC) will recur. Studies have identified multiple factors associated with recurrence, but models predicting disease relapse remain inconsistent. New analytics, like deep learning (DL), may improve prediction of EEC recurrence. Identifying patients at risk for recurrence is important to better select patients for adjuvant therapy. This study aimed to train, validate and test multimodal models of recurrence of low and high-risk EEC with DL analytics integrating genomic data and histopathologic slide images from EEC tumors. Methods Data and paraffin-embedded Hematoxylin and Eosin (H&E) slides from EEC patients were obtained from The Cancer Genome Atlas (TCGA) and stratified into low-risk, with grades 1 and 2 and stage I (recurrent: n = 16; nonrecurrent: n = 150) and high-risk, with grade 3 or stage II, III, IV (recurrent: n = 44; nonrecurrent: n = 196). Clinical and genomic data were extracted. Transcriptome data included genes, microRNA, long noncoding RNA, isoforms and pseudogenes expression. Genetic variation from whole exome sequencing included single nucleotide variation and copy number variation. In the discovery phase, informative variables for recurrence were selected with univariate ANOVA analyses (P < 0.05). Significant variables were introduced in multivariate lasso regression models. Best-performing models were validated and tested. All models' performances were measured with the area under the operating characteristic curve (AUC) and their 95 % confidence interval (CI). Data was split into 80 % for training validation and 20 % for testing with TensorFlow, a DL analytic platform. H&E slides were evaluated with the multimodal prototyping framework (MMP) pipeline using transformer architecture. Learned representations from both histopathologic and clinical-genomic models were then integrated into a multimodal transformer DL network. Results EEC recurrence models using histopathologic slides for the low-risk group had an AUC of 0.89 (95 % CI 0.81–0.96) and an AUC of 0.76 (95 % CI 0.53–0.94) for the high-risk group. The best-performing models with clinical-genomic for the low-risk group included clinical data, CNV and lncRNA. For the high-risk group, three of the four best-performing models included clinical data and CNVs. Multimodal models integrating clinical-genomic and pathologic models had a wide range of performances, with low-risk models for EEC recurrence showing an average AUC of 0.7 (range 0.5–0.94 by AUC) for the low-risk group and 0.67 on average (range 0.53–0.78 by AUC) for the high-risk group. Conclusions This study highlights the potential of using deep learning and multimodal data integration to predict EEC recurrence, particularly in low-risk groups. The inclusion of additional data and comprehensive follow-up with accurate recurrence status and timing will improve and refine the performances of these models, especially if they are to be used in clinical settings.

Multimodal deep learning models to predict endometrial cancer recurrence risk by integrating histopathology and genomic data

View Online

Abstract

Details

Metrics