Computational models developed by researchers at the Technion’s Faculty of Computer Science will make it possible to better evaluate the effectiveness of immunotherapy treatments, predict the patient’s life expectancy, improve treatments in a personalized way and warn of wrong diagnoses.
More on a similar matter
The research on the subject is led by Dr. Gil Shamai, a postdoctoral student in the Laboratory for Geometric Image Processing (GIP) led by Prof. Ron Kimmel, a faculty member in the Faculty of Computer Science and adjunct professor in the Viterbi Faculty of Electrical and Computer Engineering.
The article published in the journal Nature Communications describes a method based on technology and experiments developed by Dr. Shamai with Amir Livna, a master’s degree graduate from the laboratory, for the purpose of better evaluating the effectiveness of immunotherapy treatment in breast cancer. The researchers worked closely with Prof. Edmond Savo, director of the Institute of Pathology at the Carmel Hospital; Dr. Alexandra Kerzo, a pathology specialist at the Carmel Hospital and Dr. Antonio Polonia from the Ipatimup Pathology Laboratory in Portugal; with the support of Prof. Gil Bar Sela, director of the oncology department at the Emek Hospital.
Breast cancer has become the leading cause of death among women aged 20-59 and the most diagnosed cancer in 2021, with approximately 12% of all diagnoses worldwide. The incidence of breast cancer in the population is increasing due to population growth, environmental pollution, industrialized nutrition, climate change and more. The immune system is able, in many cases, to eliminate cancer cells in a targeted and more effective way than drug therapy.
The problem is that the cancer cells, which also developed in a long evolution, may develop mechanisms that prevent the immune system from attacking them. One of the innovative and experimental methods for treating breast cancer, in light of the success in treating other types of cancer, is immunotherapy – an approach that helps the immune system attack the cancer cells by damaging these mechanisms.
The algorithm sorted the breast cancer patients according to images of the biopsies taken from them. The algorithm placed on the left side tumors in which it detected PD-L1 expression and on the right side tumors without this expression. According to the algorithm, patients on the right side will benefit from immunotherapy, unlike patients on the left. The algorithm placed at the bottom cases where it could not decide for sure, probably because the biopsy images were not good enough. Photo: Technion spokespeople
One of the common immunotherapy treatments is based on the interaction of a protein called PD-L-1 with immune system cells. The interaction inhibits the immune system from working, and immunotherapy may prevent this interaction and thus help the immune system attack the tumor. However, the treatment is effective and approved for use only when there is an expression of 1PD-L proteins in the cancer cells themselves. Against this background, the need arose to differentiate between tumors characterized by the presence of the PD-L-1 protein in which immunotherapy is expected to help, and tumors that are not. The problem is that measuring PD-L-1 is not a simple task, and current measurement methods are expensive and imprecise.
“Unfortunately, I personally experienced the problematic nature of the existing methods during my father’s anti-cancer treatments,” says Prof. Kimmel. “These are inconsistent methods, that is – different measurements of the same phenomenon give different results. In addition, they are very expensive and are not available everywhere, and the result is that the treatments are still given in a process that involves a lot of trial and error, meaning a huge waste of time and money and of course unnecessary suffering.”
The research group developed following a request from a medical specialist who heard about the wonders of artificial intelligence and turned to Prof. Kimmel and Dr. Shamai with the question: Can artificial intelligence help in the analysis of samples (biopsy) from breast cancer for the purpose of measuring the expression of receptors and proteins?
“His proposal sounded very interesting to me and I started researching the issue,” says Dr. Shamai. “I went to expert pathologists and they explained to me that the resolution of the tissue diagnosis is not sufficient to see receptors and proteins without special molecular dyes. One of them even said it was like looking at a closed door and guessing what was in the room. This project became my research in the following years because I wanted to prove them wrong.” The research group was established a few years ago and currently employs a large group of engineers, clinicians and undergraduate, graduate and postgraduate students.
“Pathologists we spoke with said that this is an impossible task,” explains Prof. Kimmel, “because a human pathologist cannot deduce the properties of a tumor from its shape due to the huge amount of variables. The good news is that artificial intelligence technologies and especially deep learning are able to do this. The computer, unlike A pathologist, and even the most skilled pathologist, can characterize the cancer with the help of a complex analysis of its morphology.”
The first breakthrough in research occurred about three years ago and was presented in an article in the journal JAMA in which the group presented success in identifying receptors on cancer cells. The technology presented by the researchers is based on deep learning. It extracts molecular information from biopsy images that have undergone hematoxylin and eosin (H&E) staining – a common staining used to examine tissue taken in a biopsy. The staining allows the pathologist to identify in the tissue, under the microscope, the type of cancer and its degree of severity. However, the staining alone does not allow identifying proteins that are essential in determining the appropriate treatment. This is the challenge that the Technion researchers had to face.
Now, three years after the JAMA article, the researchers present success in measuring PD-L-1 protein expression on a large sample pool: biopsies from more than 3,000 breast cancer patients. According to Dr. Shamai and Prof. Kimmel, “Deep learning requires a huge amount of information, and obtaining the required type of information is not easy. For this purpose, we wrote software code to scan online sources and automatically download thousands of biopsy samples and the relevant medical information approved for research use. The help of the pathologists who signed the study was very essential.”
The computational model developed by the researchers does not “see the receptors and proteins” directly but knows how to infer their presence or absence based on the morphology of the tissue, or the “signature” that the cancer leaves in the tissue, as reflected in H&E scans. “This is such a complex relationship,” says Dr. Shamai, “that the algorithm we developed knows how to provide us with the necessary information, but we don’t really understand how it does it.”
The research is based on advanced technologies of geometric image processing. “It’s a huge and complex challenge – producing a reliable and accurate clinical picture based on partial information,” says Prof. Kimmel. “There are many involved parties here – engineers, clinicians, students and ten hospitals in Israel and abroad. The research group was required to study the clinical aspects of breast cancer and its diagnosis and to create a huge database and tools to analyze the data and create a reliable prediction. Dr. Shamai also had an enormous administrative task here, and he met it with impressive success, among other things by obtaining seven research grants from foundations that recognized the importance of the project. In fact, he established a kind of medical-technological startup here that presents dramatic achievements in the field of cancer.”
The new technology has been verified (validated) on data from 300 breast cancer patients from a clinical trial. The researchers estimate that this approach, the analysis of pathological images using artificial intelligence, will become a significant factor in cancer research, in improving diagnosis and in improving the suitability of treatment for the patient.
The research was supported by the Innovation Authority (Kamin grant) and by the National Science Foundation (the research program in personalized medicine).