Nikki Weststeijn (13/06/2023)
New method to determine the prognosis of breast cancer: AI can do it as
well as doctors
A breast tumor resection which is given as input to an AI model to determine the ratio of tumors to stroma. Photo: The Cancer Genome Atlas
One in seven women get breast cancer and they sometimes unnecessarily undergo chemotherapy, because it is hard to
determine the probability that the cancer will spread. Computer scientist Ajey Pai does research on
how we can use deep learning to estimate if the cancer is going to metastasize.
The model that Pai is working with can already perform equally well as expert
pathologists in analyzing tissue resections. “The deep learning model is going
to be more accurate and faster than traditional methods”.
Breast
cancer is the most prevalent type of cancer in the world. Luckily, it is also a
type of cancer that can often be treated successfully. Still, the treatment
process could be improved. Nowadays
doctors do not always have a way to determine whether or not the cancer is
going to spread, and chemotherapy is advised to the patient just to be sure.
Chemotherapy can, however, be a horrible experience because of its many side
effects. Hair loss, nausea and fatigue are some of the main symptoms that patients
deal with.
Because doctors cannot always make a precise diagnosis, women sometimes do extra tests in private clinics. In the Netherlands, a company called “Mammaprint” offers genetic tests that cost thousands of euros and are not covered by insurance. Computer scientist Ajey Pai does research at the Netherlands Cancer Institute and argues that it is unfair that some women can get these tests done and others cannot: “Not everybody can afford these tests”. The deep learning method that he is researching can be easily used by doctors in the hospital and will therefore allow more women to get better treatment. “Research shows that the method used by the deep learning model is a good way to determine the prognosis of different cancer types.”
The method the deep learning model will use is based on the fact that the ratio of tumorous to non-tumorous cell tissue is a good indicator of how the cancer will develop. It has been shown already that if a group of doctors is asked to estimate this ratio, the average of these estimates can indicate if the cancer is going to metastasize. The deep learning models do the same thing, but faster.
Deep learning is a type of machine learning where the model is based on the structure of the human brain and can therefore learn directly from the data, which in this case are the tissue resections. Pai’s model is taught to recognize which parts of the cell tissue are tumorous and which parts are non-tumorous, also called “stroma”. This is called supervised learning. The model can then on its own find the ratio of tumors to stroma. This part is unsupervised learning, because Pai does not teach the model how to do that.
Pai’s recent findings show that the deep learning models can perform equally well as expert pathologists in analyzing the tissue resections. The prognoses that pathologists make often differ a lot from the prognoses that other pathologists make and it turns out that the deep learning model is more often close to the consensus than individual doctors are.
The idea is that doctors can make use of the deep learning models in order to be able to treat more patients faster and in a more accurate manner. Pai: “The doctor can look at the prediction from the model, they can make an idea in their own heads and then they can accept or reject it. I’m not doing this research to replace doctors. We just want to make the lives of doctors easier.”
Even though it will be still be some time before we will see a model like this in practice, Pai believes that we will definitely employ some of them in the future. “It’s in the future but it’s very close, closer than we think. In another 30 years, I’m sure we will see some of these models come into daily practice.”
Do you wanna hear more about Ajey’s research? Check out episode 26 on Computational Cancer Research, in which Ajey was our guest !
Because doctors cannot always make a precise diagnosis, women sometimes do extra tests in private clinics. In the Netherlands, a company called “Mammaprint” offers genetic tests that cost thousands of euros and are not covered by insurance. Computer scientist Ajey Pai does research at the Netherlands Cancer Institute and argues that it is unfair that some women can get these tests done and others cannot: “Not everybody can afford these tests”. The deep learning method that he is researching can be easily used by doctors in the hospital and will therefore allow more women to get better treatment. “Research shows that the method used by the deep learning model is a good way to determine the prognosis of different cancer types.”
The method the deep learning model will use is based on the fact that the ratio of tumorous to non-tumorous cell tissue is a good indicator of how the cancer will develop. It has been shown already that if a group of doctors is asked to estimate this ratio, the average of these estimates can indicate if the cancer is going to metastasize. The deep learning models do the same thing, but faster.
Deep learning is a type of machine learning where the model is based on the structure of the human brain and can therefore learn directly from the data, which in this case are the tissue resections. Pai’s model is taught to recognize which parts of the cell tissue are tumorous and which parts are non-tumorous, also called “stroma”. This is called supervised learning. The model can then on its own find the ratio of tumors to stroma. This part is unsupervised learning, because Pai does not teach the model how to do that.
Pai’s recent findings show that the deep learning models can perform equally well as expert pathologists in analyzing the tissue resections. The prognoses that pathologists make often differ a lot from the prognoses that other pathologists make and it turns out that the deep learning model is more often close to the consensus than individual doctors are.
The idea is that doctors can make use of the deep learning models in order to be able to treat more patients faster and in a more accurate manner. Pai: “The doctor can look at the prediction from the model, they can make an idea in their own heads and then they can accept or reject it. I’m not doing this research to replace doctors. We just want to make the lives of doctors easier.”
Even though it will be still be some time before we will see a model like this in practice, Pai believes that we will definitely employ some of them in the future. “It’s in the future but it’s very close, closer than we think. In another 30 years, I’m sure we will see some of these models come into daily practice.”
Do you wanna hear more about Ajey’s research? Check out episode 26 on Computational Cancer Research, in which Ajey was our guest !