MR images contouring

The study and development of segmentation algorithms of tumor lesions on magnetic resonance images of the breast were supported by the collaboration with the Politecnico of Turin and the University of Trieste. Two master thesis projects of the Biomedical Engineering and Physics Department led to the creation of a Computer Aided Diagnosis (CAD) system for the identification and contouring of tumor targets.

The thesis works led to the implementation of various approaches for the analysis of dynamic scans with contrast medium (DCE-MRI) including Level-Set algorithms for the identification of macro-areas of interest, followed by double threshold segmentation algorithms (Seeded Region Growing) for the identification of the tumor lesion. Once the suspect regions have been identified, they are further discriminated using a heuristic approach to exclude false classifieds from subsequent analyses. On the regions identified as suspect, a parametric characterization phase follows which involves the extraction of geometric features and texture analysis parameters, useful for diagnostic purposes, starting from the automatic contouring.

The results obtained reported a number of false negatives and false positives lower than 2% and 13% respectively; for the segmentation of the target volumes, DICE index values of 98% were obtained. From the analysis of datasets acquired post-neoadjuvant therapy it has been demonstrated that CAD is able to monitor the efficacy of the therapy over time, providing not only a geometric evaluation but also a parametric characterization of the residual tumor to guarantee a more accurate evaluation. accurate.

Once the robustness of the algorithm has been verified, the next objective will be to implement it in the web-based visualization and analysis platform developed by Tecnologie Avanzate and dedicated to breast radiology (iTA Web RX).