D5.3 Deep Learning Master model and Radiomic Signatures
Deliverable 5.3, led by partner FCHAMPALIMAUD, contains the work performed by the ProCAncer-I
consortium on master models using radiomics and deep learning techniques. ’Master models’ — models
which can act as a foundation for other models — were developed for radiomics for all relevant use cases
(UC2, UC3, UC5, UC6, UC7a, UC7b and UC8) through the development of consistent and robust pipelines,
while deep learning was used only for UC1, UC2 and UC5 due to its more demanding data requirements.
Radiomics master models were developed by three partners (FCHAMPALIMAUD, FORTH and CNR), while
deep learning master models were developed and investigated by six different partners (FCHAMPALIMAUD,
CNR, FORTH, ADVANTIS, FPO, QUIBIM). Through this approach, several aspects of deep learning models
were investigated and consistent approaches and trends were identified. We finally note that the concept of
a ’master model’ is similar to that of a foundation model; in that light, this deliverable reflects that insight.
We describe the work in terms of foundation models and provide an overview of all experiments performed
to arrive at foundation models.
