Physicians routinely acquire radiologic imaging and pathological specimens for patients diagnosed with high-grade serous ovarian cancer (HGSOC). After subsequent interpretation by physicians, these data modalities inform important management decisions and prognostic evaluation.
However, can these data be mined more deeply to extract further insight about each patient’s disease? If so, how do these data complement or correlate with previously established clinical and genomic risk factors? These questions formed the basis of our study, Multimodal data integration using machine learning improves risk stratification of high-grade serous ovarian cancer.
Studying disparate clinical data modalities led to the formation of a large interdisciplinary team, spanning Computational Oncology, Diagnostic Radiology, Pathology, and Medical and Surgical Oncology. One of the best parts was running into each other all throughout the hospital!
In our study, we assembled a dataset of 444 patients with radiologic, histopathological, clinical, and genomic data and used machine learning to identify quantitative biomarkers and their relation to one another. This work derived two biomarkers from routine clinical imaging modalities and explored the interplay between them and existing clinicogenomic biomarkers. We hope our techniques can be expanded, both to further study ovarian cancer and to investigate multimodal data integration for other disease sites.
Findings and Significance
Our first major finding was on pretreatment contrast-enhanced computed tomography (CT): we found that high radiomic autocorrelation of metastatic implants in the omentum was associated with shorter overall survival. While most HGSOC radiomic studies have focused on primary adnexal lesions, we found that omental radiomic features were more strongly associated with survival outcomes. Furthermore, omental implants are much easier than adnexal lesions for radiologists to segment, a necessary preprocessing step. The omentum is also the most common site of tropism for HGSOC, making it further advantageous.
Our second major finding was on pretreatment digital histopathology: our computational model based on average tumor nuclear size and stroma morphology stratified patients by overall survival. Patients with higher mean tumor nuclear size exhibited poorer overall survival, perhaps reflecting whole-genome duplication or cellular fusion. This pipeline identified leads on some additional pathologic risk factors for HGSOC beyond stage. Though the features did not remain significantly associated with overall survival after correction for multiple testing, 20 of the 24 top features were associated with tumor nuclear size, suggesting a robust association. The stromal morphological feature is more challenging to interpret and requires further validation and future investigation to test its generalizability, especially given the variable sizes of individual tissue specimens. We hope our framework will serve as an open-source methodology to extract interpretable quantitative features from whole-slide images of tissue specimens in other cancer subtypes.
Our third major finding was that our two imaging-based models provided additive prognostic value beyond previously known clinical and genomic factors. Combining radiologic, histopathologic, and genomic risk scores improved patient stratification beyond that of any single modality. This effect held for both overall and progression-free survival. Hence, these imaging modalities at different spatial scales contain distinct prognostic information. Finally, we empirically demonstrated that our machine learning approach—which we designed to learn from cases with missing data modalities—outperformed a conventional method, which learned only from cases with all data modalities available.
In this work, we identified radiologic and histopathological biomarkers for HGSOC and developed a machine learning model to integrate them, empirically demonstrating that multi-scale data integration improves patient risk stratification. This serves as a proof of principle for such studies, though the withheld test set is relatively small. We took care to minimize the risk of overfitting (e.g., by applying uniform feature selection algorithm, withholding the test set, and minimizing the number of trainable parameters in the multimodal fusion portion of the model), and the next steps involve amalgamating a larger multi-institutional dataset to strengthen the generalizability and validation of the model. A larger dataset would also help clarify the histopathological features associated with overall survival. This work was the basis of a successful OCRA grant awarded to address these next steps. We hope that in the future, oncologists could use such models to aid in decisions around frequency of surveillance imaging, whether to administer maintenance chemotherapy, and whether to suggest clinical trials for patients at risk of suffering poor response to standard therapy.