Deep learning in cancer treatment: the algorithm that learns how to make us feel better

It is difficult to predict the course of a disease without knowing its past. For this reason, data analysis is taking on an increasingly crucial role and if talking of Big Data, the resources to manage them become absolutely necessary. At the Aviano Oncological Reference Centre (CRO), one of the main Italian oncological reference centres, a multidisciplinary team with different skills is working on this data analysis through deep learning and is getting significant results in the study of cancer.

We interviewed Maurizio Polano, a bioinformatician who deals with genomic analysis and development of artificial intelligence models in relation to precision medicine, and Roberto Ricci, a computer scientist at the Information Systems Department. 

Can we really be so precise in the diagnosis and treatment of cancer with precision medicine?

We use the data of the patient’s genetic profile to enable the doctor to make better decisions and facilitate both prevention and diagnosis. Applying scientific discoveries to therapy has always been a critical issue, and deep learning gives an important contribution to speeding up this transition because it allows to predict the evolution of the disease. However, algorithm training can be very slow and goes through repeated analysis, tests and continuous validations but once the model has been defined, it is possible to give the clinician or the researcher an easy tool to quickly support their decisions. 

Which projects are benefiting from these artificial intelligence techniques?

Several translational research projects follow the path “from bench to bedside” are working and the role of bioinformatics and artificial intelligence is becoming more and more strategic. In particular we are talking about studies on glioblastoma, radiomics and, more recently, on colorectal cancer. One of the most important studies is related to glioblastoma, one of the most devastating cancers for both patients and the carers of them. It is the most aggressive of brain cancers and only 3% of patients have a life expectancy exceeding 5 years. Therefore, it is a race against time: it is important to find better and better methods to diagnose it earlier and assess the effectiveness of the therapies. Obviously, this is not possible without having big amounts of data available.

How do you manage these large amounts of data?

A lot of computing power is needed. We have decided to use the computational resources made available by GARR over its federated cloud, in collaboration with the Institute Information System which manages the resources between the different projects. Thanks to these resources and by using GPUs (the graphics cards used for these calculations), processing (previously taking 24 hours) can be reduced to a duration of 6 hours and this allows us to test more and faster. This is decisive for training the algorithm and making it more and more reliable. 

Despite the numerous applications of AI in medicine, there are still obstacles to overcome. What are the main difficulties?

There is a problem with skills, because the bioinformatician has often a heterogeneous background, but is not a doctor, therefore communication and sharing between groups become strategic. Another  problem is related to resources: as the amount of data is growing ( the increasingly high resolution of the resonance images is an example), the computational capacity required for the analysis also increases. The possibility to use the GARR cloud makes the difference. In the field of medical research, we can often stop when faced with the impossibility of having in-house digital infrastructures, but this problem can be solved using the GARR cloud. Also, if we consider the security issues related to the sensitivity of the data we process,  we have the guarantee that the data are stored in Italy and managed on a platform run by the research community. Furthermore, we have to face the resistance to data sharing.  I wish that, in the very near future, the Research Hospitals, which are all connected to the GARR network, can adopt a data sharing approach so a model developed by an Institute can also be shared with other Institutes so that it can then be further improved and refined.