The AI model, called Crossnn, analyzes the epigenetic structure of the tumors and compares it with the detailed genetic information of over 8,000 already studied tumors.
Cancer can be treated with the help of photo: Adevărul (archive)
Magnetic resonance imaging (MRI) may indicate the presence of a brain tumor in a difficult location, but performing a brain biopsy involves significant risks for a patient, for example, presenting to the doctor with diplopia symptoms (double view).
Clinical situations such as this caused researchers from Charité – University of Berlin to explore modern diagnostic alternatives. The result: a model based on artificial intelligence, writes News.
This model to use specific features of the tumor genetic material, the epigenetic imprint, which can be obtained, for example, from the cerebrospinal fluid.
As shown in the study published on Friday, in Nature Cancer magazine, the new intelligent model allows a rapid classification of tumors, with a high degree of precision.
Currently, the diversity of tumor types significantly exceeds the number of organs of origin, reflecting the biological complexity of oncogenic processes.
Each tumor has its own particularities: histological features, growth rate, metabolic characteristics. However, tumors with similar molecular features can be grouped, and the choice of adequate therapy depends decisively on the type of tumor.
The state -of -the -art target therapies act on the specific molecular structures of the neoplastic cells or interfere with the signal pathways involved in the aberrant (pathological) cell proliferation.
Chemotherapy can be adjusted according to the tumor type. In the case of rare tumors, innovative therapies can be considered in the context of clinical studies.
“In a medical landscape marked by accelerated personalization, the precise diagnosis, made in an accredited oncological center, is essential for the effective treatment of cancer”stressed Prof. Martin E. Kreis, medical director at Charité, quoted in a statement.
Although the molecular, cellular and functional analysis of a tumor based on a tissue biopsy, it can provide the necessary information, there are cases where the tumor sample is impossible or risky.
In addition, the histological examination does not reach the accuracy offered by the new AI model.
Exploring the genome instead of standard histological analysis
For the analysis of brain tumors, a method that is not based on classic microscopic diagnosis has been validated, but on epigenetic changes, that is, how certain parts of DNA are expressed or suppressed. These are a system of “memory” cellular and controls the activation of genes.
The epigenetic information in the tumor cells presents characteristic changes, specific to each tumor type, thus allowing their precise differentiation and classification.
“Hundreds of thousands of epigenetic changes function as starting/stopping switches for specific genetic sequences. The models of these changes form a unique and unmistakable imprint“Explains Dr. Philipp Euskirchen, a researcher at the German Cancer Consortium (DKTK) and the Neuropathology Institute, the main author of the study.
In the case of brain tumors, a sample of cerebrospinal fluid is sometimes sufficient, completely avoiding the need for surgery.
In order to compare the epigenetic fingerprint of an unknown tumor with those of the thousands of tumors already cataloged, automatic learning methods are required, given the complexity and volume of information. In addition, DNA sequencing techniques are currently applied are varied, and epigenetic analyzes focused on limited genetic segments.
“We aimed to develop a model capable of classifying tumors correctly, even when epigenetic profiles are incomplete or obtained by different methods, with variable degrees of accuracy.”adds Dr. Sören Lukassen, Bioinformatician and coordinator of the Medical Group of the Berlin Institute of Health (BiH).
The proposed model uses a simplified neuronal network architecture and has been trained on a large volume of reference tumors, being subsequently validated by testing on a set of over 5,000 tumors.
“Our model allows an extremely precise diagnosis for brain tumors in 99.1% of cases and exceeds the existing solutions as accurately.”says Dr. Euskirchen.
“We also managed to train a model that distinguishes over 170 types of tumors from various organs, with an accuracy of 97.8%. This allows its application not only in the case of brain tumors, but also in cancers from other organs,” the researcher said.
An essential criterion for the validation and clinical implementation of AI models is the transparency of decisions, being imperative that the specialists can understand the mechanisms by which the algorithm reaches a certain classification.
The molecular imprint used by the model can come from both tissue and biological fluids.
In the case of brain tumors, the Neuropathology Department of Charité already offers non-invasive cerebrospinal fluid diagnosis (liquid biopsy). This avoids stressful surgery, including in complicated cases.
The patient who presented with diplopia benefited from this method.
“We analyzed the cerebrospinal fluid using nanopor sequence, an innovative, fast and efficient genetic method. The interpretation of data with our model indicated that it is a lymphoma of the central nervous system, which allowed the rapid initiation of the appropriate treatment of chemotherapy“Added Dr. Euskirchen.
The precision of the method was surprising even for the research team.
“Although the architecture of our model is considerably simpler compared to previous approaches, this allows understanding the way the predictions are generated and offers an increased level of accuracy, thus contributing to increasing the trust in diagnosis ”, says Dr. Lukassen.
In collaboration with the German Cancer Consortium (DKTK), the team intends to launch clinical studies with the AI, Crossnn model in all eight DKTK centers in Germany.
At the same time, the possibility of applying the method during surgery will be analyzed.
The purpose is to integrate this quick and accessible solution to identify the tumor type in current medical practice.
