10.08.2020 10:16
by Annemarie Post

HYBRID Publications (3): Evaluating cancer severity with the help of artificial intelligence - by Nicolò Capobianco

In June 2020, Nicolò Capobianco published the results of his collaboration with the HYBRID team in Orsay, France in the Journal of Nuclear Medicine. In the article he describes how artificial intelligence can be used to help doctors evaluate the severity of lymphoma in PET/CT images (link to publication below).

The aim of Nicolò’s project is to improve the assessment of cancer severity with PET/CT images. To achieve this goal, he focuses on developing and evaluating image analysis methods. These methods can obtain accurate measurements of the spread of the tumour and its localization in the body. Using this evaluation of disease severity, doctors can decide on the best treatment, tailor it to each patient, and monitor the progress of the disease.

In his first paper, Nicolò evaluated an image analysis method that uses artificial intelligence. The method is used to identify regions suspicious for tumour presence in PET/CT images of lymphoma patients. Moreover, he wanted to know whether this method could give an estimation of the total tumour volume, which is marker for disease severity. The method he investigated enables a rapid analysis of the images and was compared to a traditional method that requires extensive manual input by and experienced physician.



A doctor’s (scintillation) crystal ball

Medical imaging, e.g. CT and PET, offers doctors the invaluable ability to visualize and evaluate the anatomy and function of the entire human body. A big advantage of imaging compared to other techniques is that imaging is a non-invasive procedure.

Computed Tomography (CT) images are obtained by irradiating the body from multiple directions with an external X-ray source. The various organs in the body allow the radiation to pass through to a greater or lesser extent, due to their density. Therefore, by measuring the amount of radiation emerging from the body at different angles, one can obtain a three-dimensional map of the location and size of the different organs and tissues.

Positron Emission Tomography (PET) images are obtained by measuring the radiation from a molecule – called tracer – that is injected in the body. While invisible to the eye, this radiation can be detected using scintillation crystals. When these crystals are hit by radiation, they emit a sparkle of light that can be measured with electronics.  By measuring the amount of radiation that emerges from the body at different angles, one can obtain a three-dimensional map of the tracer concentration in different organs and tissues. An often-used tracer - 18F-FDG - is a molecule very similar to sugar, and thanks to PET images, one can visualize and detect tissues with a very high consumption of sugar, including tumours.

PET and CT images provide valuable complementary information, so that they are often obtained together with a single machine (PET/CT). This way, density and metabolism can be visualized at the same time, or in other words, anatomy and function.


PET/CT imaging in lymphoma

Lymphoma is a type of cancer that originates from white blood cells. There are many subtypes of lymphoma, depending on the specific cells from which the cancer develops, that differ in terms of disease progression and aggressiveness. In 2018 there were more than half a million lymphoma cases worldwide.

Physicians use 18F-FDG PET/CT images to determine the severity of lymphoma. They visualize and monitor the distribution of the tumour in the entire body and use this to decide on a more aggressive or less intense treatment.

In the past, physicians have defined rules to classify patients in different stages, based on the number and location of tumour sites. The stage, which represents the severity of the disease, determines which treatment a patient will get. This is a challenging job, because the tumour can present with highly variable shapes, sizes and locations.

More recently, researchers have shown that the measurement of the total volume of tumour in the body based on 18F-FDG PET/CT images is an indicator of the disease severity. By looking at the total tumour volume (all tumour sites in the body taken together), physicians are able to identify patients who are less likely to respond to standard treatment.


Estimating tumour burden using deep learning

The measurement of total tumour volume is not easy. A trained physician has to manually select all tumour regions in PET/CT images. This can be very time consuming, especially for patients with an advanced disease and a high number of tumour sites in the body. The physician needs to be very experienced, since he has to correctly discriminate between different tracer uptake sources (e.g. in the bowel, muscles, fat, inflamed or infected organs) and tracer uptake likely due to the tumour presence.

Nicolò has investigated whether a deep learning algorithm is also able to discriminate between nonsuspicious and suspicious tracer uptake. A deep learning algorithm has the ability to learn from a large set of examples and then evaluate new input.  Deep learning proved useful in a variety of tasks and domains, spanning from speech recognition and natural language processing to image recognition, often approaching human-level performance.

In his paper, Nicolò investigated the use of a promising image analysis method to calculate the total tumour volume in a group of about three hundred lymphoma patients.

This deep learning algorithm was trained on more than six hundred PET/CT images to identify tracer uptake sites suspicious for tumour.

The algorithm performed well, despite the PET/CT images having different quality and being acquired by different machines. Moreover, by using the algorithm to automatically identify suspicious tracer uptake regions, it proved possible to derive an estimation of the total tumour volume. Nicolò showed that this total tumour volume could predict lymphoma progression.



Nicolò concludes that the evaluated PET/CT image analysis method, based on artificial intelligence techniques leveraging deep learning, showed good ability to identify regions of tracer uptake suspicious for tumour presence. The algorithm may enable physicians to rapidly and accurately measure total tumour volume for the assessment of disease severity in lymphoma patients.


Publication: Capobianco N, Meignan MA, Cottereau AS, et al. Deep learning FDG uptake classification enables total metabolic tumor volume estimation in diffuse large B-cell lymphoma [published online ahead of print, 2020 Jun 12]. J Nucl Med. 2020; jnumed.120.242412. doi:10.2967/jnumed.120.242412

Images: 1) Nicolò Capobianco; 2) PET/CT image from iStock, not related to the publication; 3) originally published in JNM (link above) © SNMMI

Go back

HYBRID2020 Consortium |
© All Rights Reserved Imprint & Data Policy