Development and validation of cancer markers based on multiparametric MRI and machine learning

Abstract

[eng] HYPOTHESIS: The clinical management of cancer disease relies on radiological images for critical tasks like diagnosis, prognosis and assessment of response to the treatment. The current reference methods of image analysis involve visual inspection, manual annotations and a degree of subjectivity which inherently limit the characterization and evaluation of cancer for those three tasks. The development of computerized features from medical images has been an active re- search topic for years, especially within the field of neuroradiology, showing the potential of MRI for characterizing cancer disease. However, further development and validation of novel imaging markers are still needed to address the multiple clinical needs related to a complex disease. For this work, it was hypothesized that MRI-derived markers could improve over current assessments of the medical images in three identified scenarios in the management of cancer disease, specifically: • The residual tumour after surgery can be objectively quantified from MRI and be useful for estimating prognosis in patients with malignant primary brain tumours (i.e., brain glioblastoma). The time elapsed from surgery to MRI can affect or confound the estimation of residual tumour. • The use of the complete dynamic profile of perfusion MRI data can be used for the differential diagnosis of brain malignancies, and it can improve over existing methods. • The metrics derived from whole-body MRI in bone metastases associate with the re- sponse of the patients to the treatment. The MRI-derived metrics of bone metastases can capture underlying biology traits of tumours. OBJECTIVES: To test the hypotheses, three objectives were defined: • Objective 1: To extract computerized metrics from post-surgical MRI and study their prognostic value in patients with brain glioblastoma, and to explore the effect of the time elapsed from surgery to the MRI scan on the prognosis evaluation. • Objective 2: To develop and validate a method to process all time-points from dy- namic perfusion MRI for brain tumour diagnosis, differentiating between the three most common brain malignancies and comparing the performance with existing metrics. • Objective 3: To study MRI-derived data during treatment of bone metastases for the evaluation of response, exploring the combination of imaging features and their relation to histological evidence.