• Center on Health Equity and Access
  • Clinical
  • Health Care Cost
  • Health Care Delivery
  • Insurance
  • Policy
  • Technology
  • Value-Based Care

AI Accurately Predicts Therapy Outcomes in Most Patients With Ovarian Cancer


An artificial intelligence (AI)–based prediction model correctly predicted outcomes for 78% of patients with high-grade serous ovarian carcinoma, with an accuracy of 80%.

An artificial intelligence (AI)–based prediction model was able to correctly predict therapy outcomes in 78% of patients with high-grade serous ovarian carcinoma (HGSOC) in a study published in Nature Communications.

The model was found to be 80% accurate, which the researchers said is “significantly better than current clinical methods,” and therapy outcomes were determined by the volumetric reduction of ovarian tumor lesions.

AI data technology | Image credit: NicoElNino – stock.adobe.com

AI data technology | Image credit: NicoElNino – stock.adobe.com

HGSOC is a complex and varied disease often detected at an advanced stage, and predicting how patients will respond to neoadjuvant chemotherapy (NACT) and understanding the key factors influencing response is challenging due to the diverse nature of the HGSOC. To potentially help strengthen these predictions, researchers developed an Integrated Radiogenomics for Ovarian Neoadjuvant therapy (IRON) framework that considers several factors, including initial clinical information, blood-based markers, and radiomic biomarkers gathered from both primary and metastatic lesions.

The AI-based model combines these different types of data to predict changes in total disease volume based on information collected at the time of diagnosis; it was tested on 72 patients with HGSOC. IRON’s accuracy was then tested on 2 additional groups: an internal cohort of 20 patients and an external cohort of 42 patients.

In the external group, IRON performed significantly better, reducing prediction errors by 8% compared with a model based solely on clinical information. This integrated model achieved an area under the curve of 0.78 for response evaluation criteria in solid tumors (RECIST 1.1) classification, while the clinical model reached only 0.47. According to the researchers, these findings underscore the importance of including radiomics data in models assessing treatment response and suggest new approaches for designing biomarker-based clinical trials for NACT in HGSOC.

The response patterns to NACT were found to be heterogeneous. The researchers noted that all primary and metastatic lesions identified in CT scans before and after NACT were carefully segmented and labeled. Notably, tumors in the omentum and pelvic or ovarian regions constituted most of the disease burden initially and were the most common tumor locations, and significant variations in treatment response were observed among patients at the same anatomic sites. Particularly, the omental disease exhibited a notably better response compared with pelvic disease.

Despite differences in response at specific locations, the application of RECIST 1.1 strongly correlated with the overall change in total volume. Interestingly, among responders assessed by RECIST, the baseline volume of omental disease was higher compared with nonresponders, emphasizing the importance of detailed volumetric data in predicting NACT response.

“Taken together, these detailed volumetric data indicate that multivariable predictors are required to predict response to NACT rather than simple knowledge about disease burden and its anatomic distribution,” the researchers said.

The study also revealed that circulating tumor DNA (ctDNA) and cancer antigen 125 (CA-125) correlated with different types of disease burden for patients with HGSOC. In the training and hold-out validation cohorts, the researchers evaluated ctDNA at baseline for all patients. They compared various measures, including TP53 mutant allele fraction (MAF), trimmed median absolute deviation from copy number neutrality (t-MAD) based on shallow whole-genome sequencing, and computed haploid genome equivalents per mL (hGE/mL). All of these measures showed a high correlation, with TP53 MAF being selected for further analyses. The total disease burden at baseline—including total volume, number of lesions, and summed RECIST 1.1 diameters—exhibited significant correlations with both CA-125 and TP53 MAF.

While baseline CA-125 and TP53 MAF did not correlate with treatment response, they did show a significant positive correlation with the summed RECIST 1.1 diameters post chemotherapy. Additionally, baseline CA-125 correlated with omental and pelvic/ovarian disease volume measured before chemotherapy, while baseline TP53 MAF correlated significantly with pelvic/ovarian disease volume measured both before and after chemotherapy. The findings suggest that a high baseline TP53 MAF could be a specific indicator for a high disease burden in the ovaries or pelvis, which tends to show a poorer response.

The study also explored the correlation between radiomics features and clinical as well as biological characteristics in the context of HGSOC. Various collections of radiomics features were defined to capture the radiological complexity of the disease, including volumes, lesion numbers, shape features, first-order histogram statistics, texture features, intra-lesion heterogeneity, and peripheral context of lesions.

The researchers identified 6 distinct clusters of imaging features, with each cluster associated with specific biological or clinical factors. For example, cluster 1 correlated with baseline CA-125 levels and contained mainly lesion volume metrics, which is consistent with previous findings. Additionally, clusters associated with ctDNA features were dominated by features quantifying lesion heterogeneity and context, and clusters 4 and 2 were highly correlated with disease stage and patient age, respectively.

According to the researchers, these findings suggest that information from global biomarkers like disease stage, CA-125, or ctDNA can also be captured in multilesion radiomic features, providing insights into the extent, spread, heterogeneity, and context of the disease. Importantly, some imaging features within certain clusters showed significant correlations with volumetric treatment response, suggesting the potential of radiomics in contributing unique information to integrated radiogenomic predictive models.

The systematic integration of standard-of-care biomarkers on multiple scales offers crucial predictive capabilities and valuable insights into the intricate spatial configuration of the disease. The researchers emphasized that the potential clinical advancement of IRON could significantly impact patient stratification in both clinical and experimental settings, potentially preventing delays in surgery for patients unlikely to respond to chemotherapy and facilitating the design of new-generation clinical trials for HGSOC with efficient and accelerated end points, enhancing the discovery of novel therapies.

“The field of radiomics has grown exponentially in recent years, showing great promise across tumour sites and endpoints,” the researchers wrote. “At the same time, radiomics has been criticised for lack of robustness and reproducibility, as well as lack of biological interpretability. Our study shows that both problems can be overcome by the right design choices.”


Crispin-Ortuzar M, Woitek R, Reinius MAV, et al. Integrated radiogenomics models predict response to neoadjuvant chemotherapy in high grade serous ovarian cancer. Nat Commun. 2023;14(1):6756. doi:10.1038/s41467-023-41820-7

Related Videos
Dr Guru Sonpavde
Dr Jeffrey Sippel
Dr. Jeffrey Sippel
Dr. Jeffrey Sippel
Takiyah Durham, MBA, and Margaret Larkins-Pettigrew, MD
Related Content
© 2023 MJH Life Sciences
All rights reserved.