Penn Medicine, Intel Leverage Federated Learning for Tumor Detection

December 07, 2022 - Researchers from the Perelman School of Medicine at the University of Pennsylvania (Penn Medicine) and Intel Labs have announced the results of a joint study that leveraged federated learning (FL) to help clinicians and researchers identify malignant brain tumors in glioblastoma patients.

The University of Texas MD Anderson Cancer Center reports that glioblastoma is the most common primary brain cancer, with around 12,000 cases diagnosed in the US annually. It is an aggressive type of cancer, with a median length of survival after a diagnosis of 15 to 18 months. Glioblastoma’s five-year survival rate is also about 10 percent, which reflects some of the challenges in treating the disease.

MD Anderson notes that treating glioblastoma can be difficult because of the nature of brain cancers. When removing tumors from other parts of the body, surgeons typically also remove a small section of healthy tissue around the tumor, which helps remove cancer cells not visible to the naked eye. With brain tumors, removing this margin of healthy tissue can be challenging because the tissue surrounding a tumor may control essential functions like speech and movement.

Detecting tumors and determining the operable region around a tumor can help address these challenges, according to a press release published alongside the study. In 2020, Penn Medicine and Intel announced an agreement to collaborate and use FL to improve tumor detection and outcomes for glioblastoma. The study aimed to further that goal by improving the detection and identification of tumor boundaries.

By delineating a tumor’s boundary, clinicians can gain insights into where healthy tissue and the tumor meet and determine what section of tissue to remove for best results during surgery. However, defining a tumor’s boundaries can be complex depending on various factors. Further, finding a glioblastoma or other brain cancer’s boundaries can be even more difficult based on the rarity of the cancer type and the availability of imaging data.

Data privacy and sharing concerns provide a further barrier to accessing these data, but FL seeks to address the concerns by allowing each participating health system’s data to remain on-site instead of being transferred to a central repository for research. From there, algorithms are trained across all the decentralized devices or servers. Once trained on the data, the algorithms can be shared without the risk of sharing any patient data.

The press release states that the study, which used data from 71 institutions across six continents, is a proof-of-concept for using FL in medical research.

Each institution used Intel’s federated learning hardware and software to help radiologists determine the boundary of a tumor and improve the identification of the operable region of tumors. The researchers annotated the patient data and used OpenFL, an open-source framework for training machine-learning algorithms, to run the federated training.

The model was trained on 3.7 million images from 6,314 glioblastoma patients, which the study states is the largest dataset of its kind in literature to date. Following training, the research team tested the model’s ability to automatically detect tumor boundaries against that of a publicly trained model. They found that the FL model was linked to a 33-percent delineation improvement for surgically targetable tumors.

“In this study, federated learning shows its potential as a paradigm shift in securing multi-institutional collaborations by enabling access to the largest and most diverse dataset of glioblastoma patients ever considered in the literature, while all data are retained within each institution at all times,” said Spyridon Bakas, PhD, senior author of the study and assistant professor of pathology and laboratory medicine and radiology at the Perelman School of Medicine, in the press release. “The more data we can feed into machine learning models, the more accurate they become, which in turn can improve our ability to understand and treat even rare diseases, such as glioblastoma.”