Artificial Intelligence in Oncology - Supporting scientific research
UMC Utrecht
In short, Kenneth Gilhuijss' research entails the following:
'Multiple Myeloma (MM) is a rare hematological malignancy with age-adjusted incidence of six per 100.000 per year in Europe and the USA. The most feared manifestation - in addition to anemia, renal dysfunction and hypercalcaemia - is extensive skeletal damage; Holes develop in the skeleton (osteolytic lesions) and bone decalcification lead to vertebral collapses.
No methods are currently available to predict at which vertebral levels such collapses will occur in individual patients with sufficiently high specificity. If such methods are available, they could serve to triage patients to preventive interventions at the earliest sign of spinal weakening, preferably before vertebral collapses occur.
The aim of this project is to make patient-specific vertebral-level oriented predictions of an imminent and clinically significant Vertebral Collapse Fracture (VCF). For this purpose, we pursue Artificial Intelligence to integrate baseline and follow-up CT scans and clinicopathological features. We expect that this will reduce the risk of (further) vertebral collapse, thus improving the quality of life and survival of MM patients.
PreVeCAIMM is a collaboration between the Image Sciences Institute, the departments of Hematology, Radiology, and Orthopedic surgery of the University Medical Center Utrecht.'
Deep learning was used to perform automated vertebrae identification and segmentation of CT scans of multiple myeloma patients. The performance of a state-of-the-art method contained inconsistencies that are not acceptable in the clinic and led to difficulties tracking the vertebrae in longitudinal scans. To mitigate this problem, a module was developed that automatically detects highly uncertain regions based on a-prior anatomical knowledge of the spine. The module automatically corrected the majority of such errors. Other regions considered suspicious with less certainty were automatically marked for human inspection. These areas were semi-automatically corrected using minimal human intervention. This human-in-the-loop approach to vertebrae segmentation significantly improved segmentation performance and longitudinal tracking of the vertebrae, while minimizing the workload to correct these segmentations.
Radiomics was used to extract first-order, 3D shape and texture features from the segmented vertebrae of multiple myeloma patients aiming to analyze changes over time. In addition, to complement radiomics, differential geometry was used to quantify the deformation of vertebrae over time.
A comprehensive PreVeCAIMM database was constructed that contains all consecutive patients with multiple myeloma presented at UMC Utrecht between 2005 and 2022. The first Computed-Tomography (CT) scan was typically made at the time of diagnosis, and subsequent follow-up scans when symptoms worsened. In total, 1.740 scans from 528 patients were included. The number of scans acquired per patient ranged from one to twenty-nine.
The longitudinal whole-body CT scans contain extensive imaging information. To isolate the information required for predictive analysis, a vertebral tracking and segmentation module was developed based on convolutional neural networks. The vertebral tracking module was able to locate the center points of the vertebrae and their correct vertebral labels. This enables assessment of disease progression in individual vertebrae over time. A segmentation module was subsequently developed to automatically label the voxels in each individual vertebra, thus allowing automated analysis of bone density, shape and lytic lesions over time. These characteristics are predictors for vertebral compression fractures. The vertebral tracking and segmentation module was trained and validated on an independent open-source dataset (VerSe). Testing on the PreVeCAIMM database shows promising results on generalizability of this module.