Harvard Medical School Develops TOAD, an AI-Based Diagnostic System for Cancers of Unknown Primary Origin
In most cancers, the primary site of origin of the tumor is easily discernable but in some tumors, the primary site of tumor origin is not easily identifiable. In those cases, it is a diagnostic challenge to locate the original site of the tumor and differentiate it as a primary or a metastatic tumor. Generally, a patient has to go through a series of additional tests to reach a conclusive diagnosis in those cases. This also translates to additional costs for the patients.
Moreover, most cancer treatments are targeted at the primary site of tumor origin so, in tumors in which the site of origin is unknown, there is a poor prognosis and poor survival rates.
Recently, Harvard Medical School researchers at the Mahmoud lab at Brigham and women’s hospital presented a solution to this problem of locating the origin of the tumor by developing an artificial intelligence-based diagnostic system. This system utilizes the same histology slides that are used for conventional diagnostic purposes. It identifies the primary site for tumor origin as well as provides differential diagnoses for the patients.
According to one of the researchers, this system is particularly useful for such complicated cases in which there is no clear diagnosis and it will also be cost-effective. Hence, its utility in low financial resource settings will also be high.
This artificial intelligence-based system is called tumor origin assessment via deep learning (TOAD). This system differentiates between primary and metastatic tumors and identifies their site of origin. This system was trained using 22000 histology whole slide images of cancer cases and it was tested on 6500 cancer cases with a known site of origin. The system was tested for tumors with known primary sites as well as for tumors for which there was no known primary site.
In the cancer cases of known primary origins, this system identified the origin of the tumor correctly 83 percent of the time and listed the diagnosis in the top three predictions 96 percent of the time.
In 317 patients with unknown primary origins of the tumor, a differential diagnosis was established and it was observed that the TOAD diagnosis agreed with that of pathologists 61 percent of the time. And researchers also observed that the pathologists’ diagnosis was among the top three diagnoses generated by TOAD 82 percent of the time.
Although results produced by TOAD are similar to results produced by genomic data-based tumor origin prediction and genomic data system is an alternative to this system, the genomic data system is rarely used in low resource settings.
Currently, the researchers are working on a histology-based system to improve it by including more cases and testing its effectiveness to improve diagnosis and prognosis.
According to researchers, this whole slide system can improve diagnosis by reducing the costs of additional tests and save time by removing the need for additional tests. Researchers think that this system can be used to improve the diagnostic process in low resource settings where the pathology expertise is not available and they believe that it is only the start of the artificial intelligence assisted cancer origin prediction and this area presents immense potential for future research and work.
Harvard medical school, Pinpointing Cancer’s Epicenter, Accessed May 7, 2021,
Nature, AI-based pathology predicts origins for cancers of unknown primary, Accessed May 7, 2021, https://www.nature.com/articles/s41586-021-03512-4