Point/Counter

Will advances in AI lead to more effective screening practices for ovarian cancer?

Click here to read the Cover Story, “AI an emerging tool, not substitute, for oncologists.”

POINT

Yes.

Screening techniques for ovarian cancer are nonexistent. Studies have looked at combinations of blood markers and ultrasounds, but they did not show a survival improvement. Ultimately, if a patient undergoes screening, it should lead to some improvement in outcome or prevention of the cancer altogether.

John M. Nakayama, MD
John M. Nakayama

AI can help us in this area because cancer results from a combination of genetic and environmental factors. We can try to account for more factors using AI, such as other genetic markers, epigenetic markers, changes in lipids, earlier CT or ultrasound findings that could be predictive. Putting those together would generate huge amounts of data. In fact, in a clinical commentary published in Gynecologic Oncology, McDonald and colleagues reported that the amount of genomic data alone doubles every 6 to 7 months and is predicted to exceed 40 exabytes a year within the next decade.

Given the overwhelming amount of complex data, the only way to get through it and make correlations among it all is with the help of machines. AI has generally worked in finding correlations among data.

Two studies explored this issue. One looked at the metabolome, meaning researchers used machine learning to look at different lab values and found they were able to predict ovarian cancer. The use of a support vector machine-based learning algorithm to identify 16 diagnostic metabolites detected early-stage ovarian cancer with 100% accuracy.

The other study screened for microRNAs and found just a handful of them were useful in predicting if a patient had ovarian cancer, borderline cancer or if it was benign. The researchers developed a micro-RNA algorithm for diagnosis that outperformed CA125 screening and appeared accurate regardless of patient age, histology or stage. The network also had 100% specificity for epithelial ovarian cancer when tested in a group of 454 patients with various diagnoses.

This work is preliminary and needs to be validated in a prospective manner, but these are the kinds of changes that have to happen in order for us to get effective screening for ovarian cancer. They have to be done algorithmically.

Deep learning can help with anything as long as it’s coded. The data are out there. It might be cost prohibitive to label it in some situations, but is it possible to find it and turn it into a screening method? The answer is yes.

References:

Gaul DA, et al. Sci Rep. 2015;doi:10.1038/srep16351.

McDonald JF. Gynecol Oncol. 2018;doi:10.1016/j.ygyno.2018.03.053.

Elias KM, et al. Elife. 2017;doi:10.7554/eLife.28932.

John M. Nakayama, MD, is an obstetrician-gynecologist at University Hospital Cleveland Medical Center. He can be reached at john.nakayama@uhhospitals.org. Disclosure: Nakayama reports no relevant financial disclosures.

COUNTER

No.

Any time you have AI, you need big data and you need algorithms to process and analyze those data.

Joshua P. Kesterson, MD
Joshua P. Kesterson

With ovarian cancer, several factors make this very challenging. First, this is a relatively rare entity, so accumulating large amounts of data is problematic. Second, ovarian cancer has significant heterogeneity, including different histologic subtypes, pathologies, etiologies and natural histories, making large data sets even more difficult to accumulate. And third, each hospital has its own electronic medical records, each effectively existing within its own silo, without crosstalk between institutions and data sources. All three of these factors speak to the difficulties inherent in generating machine learning to help create a screening system for ovarian cancer.

Further, for ovarian cancer, the natural history of the disease process has not been fully elucidated. We would, rather simplistically, like to assume things progress in a stepwise fashion, going from stage I to stage II to stage III to stage IV, but that’s not really what we see clinically. We see patients present with advanced-stage disease, involving the abdomen, peritoneal surface and/or lungs. We don’t know in ovarian cancer if a patient has a detectable premalignant phase, or the length of time of that phase if it does exist, and we don’t know if cancer progresses in a stepwise fashion from local anatomic site to distant anatomic site. Although tumor markers, in combination with other screening modalities, have been implemented with limited success previously to detect early-stage disease, these approaches are imperfect, lacking necessary sensitivity and specificity. At present, the process of tumorigenesis and metastasis is uncertain. Without knowing the natural history of the disease, it is difficult to create an algorithm and tell it what precursors or data points to look for.

In summary, in the confines of a limited and variable data set that can be created of patients with ovarian cancer, and lacking a more complete understanding of the biologic underpinnings of the disease, AI may not be the ideal means of improving the outcomes of those with ovarian cancer via screening.

Joshua P. Kesterson, MD, is an obstetrician-gynecologist and associate professor in the department of obstetrics and gynecology at Pennsylvania State University Cancer Institute. He can be reached at jpk14@psu.edu. Disclosure: Kesterson reports no relevant financial disclosures.

Click here to read the Cover Story, “AI an emerging tool, not substitute, for oncologists.”

POINT

Yes.

Screening techniques for ovarian cancer are nonexistent. Studies have looked at combinations of blood markers and ultrasounds, but they did not show a survival improvement. Ultimately, if a patient undergoes screening, it should lead to some improvement in outcome or prevention of the cancer altogether.

John M. Nakayama, MD
John M. Nakayama

AI can help us in this area because cancer results from a combination of genetic and environmental factors. We can try to account for more factors using AI, such as other genetic markers, epigenetic markers, changes in lipids, earlier CT or ultrasound findings that could be predictive. Putting those together would generate huge amounts of data. In fact, in a clinical commentary published in Gynecologic Oncology, McDonald and colleagues reported that the amount of genomic data alone doubles every 6 to 7 months and is predicted to exceed 40 exabytes a year within the next decade.

Given the overwhelming amount of complex data, the only way to get through it and make correlations among it all is with the help of machines. AI has generally worked in finding correlations among data.

Two studies explored this issue. One looked at the metabolome, meaning researchers used machine learning to look at different lab values and found they were able to predict ovarian cancer. The use of a support vector machine-based learning algorithm to identify 16 diagnostic metabolites detected early-stage ovarian cancer with 100% accuracy.

The other study screened for microRNAs and found just a handful of them were useful in predicting if a patient had ovarian cancer, borderline cancer or if it was benign. The researchers developed a micro-RNA algorithm for diagnosis that outperformed CA125 screening and appeared accurate regardless of patient age, histology or stage. The network also had 100% specificity for epithelial ovarian cancer when tested in a group of 454 patients with various diagnoses.

This work is preliminary and needs to be validated in a prospective manner, but these are the kinds of changes that have to happen in order for us to get effective screening for ovarian cancer. They have to be done algorithmically.

Deep learning can help with anything as long as it’s coded. The data are out there. It might be cost prohibitive to label it in some situations, but is it possible to find it and turn it into a screening method? The answer is yes.

References:

Gaul DA, et al. Sci Rep. 2015;doi:10.1038/srep16351.

McDonald JF. Gynecol Oncol. 2018;doi:10.1016/j.ygyno.2018.03.053.

Elias KM, et al. Elife. 2017;doi:10.7554/eLife.28932.

John M. Nakayama, MD, is an obstetrician-gynecologist at University Hospital Cleveland Medical Center. He can be reached at john.nakayama@uhhospitals.org. Disclosure: Nakayama reports no relevant financial disclosures.

PAGE BREAK

COUNTER

No.

Any time you have AI, you need big data and you need algorithms to process and analyze those data.

Joshua P. Kesterson, MD
Joshua P. Kesterson

With ovarian cancer, several factors make this very challenging. First, this is a relatively rare entity, so accumulating large amounts of data is problematic. Second, ovarian cancer has significant heterogeneity, including different histologic subtypes, pathologies, etiologies and natural histories, making large data sets even more difficult to accumulate. And third, each hospital has its own electronic medical records, each effectively existing within its own silo, without crosstalk between institutions and data sources. All three of these factors speak to the difficulties inherent in generating machine learning to help create a screening system for ovarian cancer.

Further, for ovarian cancer, the natural history of the disease process has not been fully elucidated. We would, rather simplistically, like to assume things progress in a stepwise fashion, going from stage I to stage II to stage III to stage IV, but that’s not really what we see clinically. We see patients present with advanced-stage disease, involving the abdomen, peritoneal surface and/or lungs. We don’t know in ovarian cancer if a patient has a detectable premalignant phase, or the length of time of that phase if it does exist, and we don’t know if cancer progresses in a stepwise fashion from local anatomic site to distant anatomic site. Although tumor markers, in combination with other screening modalities, have been implemented with limited success previously to detect early-stage disease, these approaches are imperfect, lacking necessary sensitivity and specificity. At present, the process of tumorigenesis and metastasis is uncertain. Without knowing the natural history of the disease, it is difficult to create an algorithm and tell it what precursors or data points to look for.

In summary, in the confines of a limited and variable data set that can be created of patients with ovarian cancer, and lacking a more complete understanding of the biologic underpinnings of the disease, AI may not be the ideal means of improving the outcomes of those with ovarian cancer via screening.

Joshua P. Kesterson, MD, is an obstetrician-gynecologist and associate professor in the department of obstetrics and gynecology at Pennsylvania State University Cancer Institute. He can be reached at jpk14@psu.edu. Disclosure: Kesterson reports no relevant financial disclosures.