ARtificial Intelligence for Gross Tumour vOlume Segmentation

Lung cancer (LC) is the single leading cancer cause of death worldwide (age-standardized rate of 18.5 per 100,000 population), outstripping the mortality from cancers of the breast, gastro-intestinal tract and reproductive organs. Radiotherapy (RT), often in combination with other treatments, has an essential role in managing LC. An essential step in the RT process is to draw the outline of the Gross Tumor Volume (GTV) in the lung on axial computed tomography (CT) scans. The step is required for precisely directing tumoricidal radiation to the target, and simultaneously avoiding irradiation of adjacent healthy tissue as much as reasonably achievable.

However, tumor outlining by hand consumes a large amount of expert physician time, and has demonstrably high levels of inter- and intra-observer variability. Part of a clinical solution would require validated automated systems that work well for complex GTVs in a wide variety of clinical settings. In recent times, a subclass of artificial intelligence known as deep learning neural networks (DLNNs) has shown promising potential to assist clinicians for such image processing tasks. The immense appeal of DLNN-based tools, if they can be safely shown to add value into radiotherapy clinical workflow, is easily understandable - these have the potential to significantly boost the productivity of clinicians by automating a portion of labor-intensive work.

In respect to LC, models trained on selective data from few institutions are the norm. What the field lacks is not simply large sample size, but sufficient diversity and heterogeneity of subjects to represent the real world, and the means to train a DLNN on such a population. That such a population exists among all the RT clinics around the world is indisputable, however the question is how do we utilize data from all over the world for such a purpose.

"Federated Learning" very clearly addresses this by side-stepping a few of the administrative complication of transferring individual-patient level data across national borders. Federated learning is an implementation of the Personal Health Train (PHT) paradigm, where we send research questions to each other in the form of software and exchange anonymous statistical results (such as a DLNN model) instead of sending patient data around. Hence PHT addresses two of the major challenges of using large-scale cancer data at a single stroke: (a) using data for a good purpose in spite of the geographic dispersion of oncology data, and (b) reducing privacy concerns associated sharing of private patient data across borders.

Objective