OK Google: Scan This Slide
An algorithm trained to detect breast cancer tumors could be used as a “spell check” for pathologists
Luke Turner |
Google is the most frequently used search engine around the world – but what if its power of investigation could be used to search for tumors? The artificial intelligence (AI) team at Google have taken a step toward realizing this ambition by developing an algorithm that could act as a “spell check” for pathologists. Known as LYmph Node Assistant (LYNA), the tool can scan digital images of breast cancer patients’ lymph nodes to detect how much of the cancer has spread beyond the breasts. Yun Liu of the Google AI team hopes that the method will allow pathologists to work more efficiently and accurately.
“We trained the algorithm to identify metastatic breast cancer in lymph node specimens. As it saw more and more examples, the algorithm gradually learned to distinguish tumor from non-tumor, to the point where it is more than 99 percent accurate on image patches,” Liu says. This figure should be interpreted with a degree of caution, because – although it refers to the ability to identify whether or not an image contains cancer – most of the lymph node images shown to the algorithm did not contain cancer. Despite this, the algorithm was able to detect the exact location of 91 percent of tumors from the CAMELYON dataset (when allowed one false negative per slide). Considering that a pathologist detected 72 percent of tumor foci in the same 130 slides over a 30-hour period, the algorithm’s accuracy is striking.
When LYNA was tested with pathologists to digitally review lymph node slides, the team found that those given the tool performed better than either the algorithm or the pathologists alone. Liu underlines these encouraging results. “The algorithm halved the time taken to review micrometastasis from about two minutes to one minute per slide, and also improved the micrometastasis detection sensitivity from 83 to 91 percent.” In addition, pathologists reported that cases became easier to review with LYNA’s help, offering hope that the algorithm could be used to assist overworked pathologists.
“Our ambition is to reduce the amount of time it takes to complete tedious tasks and improve accuracy when detecting lesions that are easy to miss,” Liu continues. “We hope that technologies such as this will free up pathologists to focus on the more complex, challenging, or rare cases.” The AI team at Google hope to see future work focusing on human-model interaction and are currently investigating the benefits and pitfalls associated with their algorithm’s use in clinical workflows.
- Y Liu et al., “Artificial intelligence-based breast cancer nodal metastasis detection”, Arch Pathol Lab Med, [Epub ahead of print] (2018). PMID: 30295070.
- DF Steiner et al., “Impact of deep learning assistance on the histopathologic review of lymph nodes for setastatic breast cancer”, Am J Surg Pathol, [Epub ahead of print] (2018). PMID: 30312179.