Abstract
Multimodal models for predicting chemotherapy response in high-grade serous ovarian carcinoma: Integration of deep learning models of imaging and genomic data
Gynecologic oncology, Vol.200(Supplement 1), p.299
09/2025
DOI: 10.1016/j.ygyno.2025.04.420
Abstract
Objectives
Ovarian cancer remains a leading cause of mortality from gynecologic cancer. Patients with high-grade serous ovarian cancer (HGSOC) who do not respond to initial chemotherapy have poor outcomes. Validated and accurate prediction models that will discriminate the patients at risk of failing initial systemic treatment would be valuable to individualize and strategize alternative novel therapies. Previous studies have shown a fair to moderate prediction model using histopathology slides alone to predict chemotherapy response. We hypothesize that deep learning (DL) multimodal prediction models for HGSOC chemo response, which integrates genomic and histopathologic DL models, are more robust and accurate than previous DL models using histopathology alone.
Methods
Digital whole slide images and associated clinical-genomic information from The Cancer Genome Atlas (TCGA) HGSOC dataset were downloaded. We included 108H&E stained histopathology slides from the TCGA HGSOC dataset and 91 chemo response information. Genomic data included clinical data, gene and gene isoform expressions, single nucleotide variation, copy number variation (CNV) and DNA methylation. Pre-processing pre-trained self-supervised learned histology extraction and transformer architecture modeling were performed using the MultiModal Prototyping framework (MMP) pipeline. Model performance was assessed with C-statistics using the area under the receiver operating characteristic curve (AUC) and the area under the precision-recall curve (AUPRC) and their 95 % confidence intervals (CI).
Results
The best-performing models for chemo response prediction with clinical-genomic data included a model with only clinical data (AUC 0.79 [0.71–0.88]), CNV (AUC 0.75 [0.62–0.88]) and methylation data (AUC 0.65 [0.52–0.68]). Models trained with pathology slides and validated with cross-validation (5 k-fold) predicted chemo response with an AUC of 0.75 and AUPRC of 0.5. Multimodal models of chemo-response, including the best-performing clinical-genomic and pathologic models, did not improve overall performance, with AUCs up to 0.63 and a range of 0.48–0.63.
Conclusions
Multimodal models had a wide range of results measured in AUC. Integration of images and genomic data did not improve the prediction of chemo response performance. This is most likely due to the small sample size. Multi-institutional collaboration and larger datasets would overcome these challenges. In addition, capturing and optimizing the diagnosis of the outcome of interest may improve results as chemo response is not always specifically reported in TGCA dataset.
Details
- Title: Subtitle
- Multimodal models for predicting chemotherapy response in high-grade serous ovarian carcinoma: Integration of deep learning models of imaging and genomic data
- Creators
- Katherine Sawaya - University of IowaAndrew Polio - University of IowaMichael Goodheart - University of IowaVincent Wagner - University of IowaDavid Bender - University of IowaJesus Gonzalez Bosquet - University of Iowa, Obstetrics and Gynecology
- Resource Type
- Abstract
- Publication Details
- Gynecologic oncology, Vol.200(Supplement 1), p.299
- DOI
- 10.1016/j.ygyno.2025.04.420
- ISSN
- 0090-8258
- eISSN
- 1095-6859
- Publisher
- ACADEMIC PRESS INC ELSEVIER SCIENCE
- Language
- English
- Date published
- 09/2025
- Academic Unit
- Obstetrics and Gynecology
- Record Identifier
- 9984969244202771
Metrics
8 Record Views