Nature: Artificial intelligence AI predicts cancer origin and improves the diagnosis of complex metastatic cancer

As an emerging disruptive technology, artificial intelligence is gradually releasing the huge energy accumulated by scientific and technological revolution and industrial transformation, and profoundly changing the way of human production and lifestyle
.
It can be said that artificial intelligence has had a significant and far-reaching impact
on economic development and social progress.

At present, artificial intelligence has shown its skills in mobile phone AI, face voice recognition, Go and other fields, and is constantly expanding its application fields
.
It is worth noting that "AI+ medical care" has always been highly expected, which can reduce the medical burden while reducing the occurrence
of misdiagnosis and missed diagnosis.

In May 2021, the Faisal Mahmood team at Harvard Medical School published a research paper
in Nature titled: AI-based pathology predicts origins for cancers of unknown primary.

The research team developed an artificial intelligence (AI) system that uses conventional histological sections to accurately find the origin of metastatic tumors and make differential diagnoses
of Cancer of Unknown Primary (CUP).
The AI system is able to improve the diagnosis of patients with complex metastatic cancer, especially in under-resourced settings
.

In 1-2% of cancer cases, it is not possible to determine the site where the tumor first occurs, that is, the primary focus
.
Because many modern cancer treatments target primary tumors, Cancer of Unknown Elementary (CUP) has a poor prognosis, with an average overall survival of only 2.
7-16 months
.

To obtain a more specific diagnosis, patients often have to undergo extensive diagnostic tests, which may include additional laboratory tests, biopsies, and endoscopy, among others, which can lead to delays in treatment and are detrimental
to patient survival.

Almost every patient diagnosed with cancer has a tissue section, which has been the diagnostic standard
for more than a hundred years.
Artificial intelligence (AI) combined with this available and common data can greatly improve these complex cases
, which often require a lot of manual diagnosis.

In this study, the research team developed a deep learning-based algorithm and named it "Deep Learning to Assess Tumor Origin" (TOAD), which can identify tumors as primary or metastatic tumors and predict their primary foci
.

The research team trained the AI system using billion-pixel pathology full-slice images of more than 22,000 cancer cases, then tested it in about 6,500 known primary cancer cases and analyzed increasingly complex metastatic cancers to build the AI model's analytical capabilities
on cancers of unknown primary (CUP).

For tumors with known primary lesions, the prediction accuracy of the AI model is 83%, and the prediction accuracy of the Top3 is 96%.

The research team then tested the AI model in 317 cancers of unknown primary (CUP) and found that the AI model's diagnosis was 63% consistent with pathologists and 82%
with Top3 diagnosis.

The AI model is roughly as effective as
several recent studies that use genomic data to predict tumor origin.
Although genomic-based AI technology offers an alternative to diagnosis, patients often do not have genetic testing data, especially in low-resource settings
.

The research team said that the AI model can reduce the number of auxiliary examinations that need to be taken, reduce additional tissue sampling, reduce the total time required for patient diagnosis, and speed up diagnosis and follow-up treatment
.
In addition, the Top3 prediction results can guide pathology to narrow down the scope and make follow-up operations
simpler.

This is the first step in AI-assisted cancer origin prediction using whole-tissue section images and is an exciting field with the potential to
standardize and improve the diagnostic process.
The research team hopes to continue training this histology-based AI model in more cases and participate in clinical trials to determine whether it can improve diagnostic capabilities and patient outcomes
.

Links to papers: