Tumor response algorithm that corrects for anatomical changes and robustly identifies local changes in serial PET images through a clustering algorithm.
- Robust to noise and outliers and can tolerate imperfections in data acquisition.
- Could be applicable to many other imaging modalities in radiation therapy.
- More accurately assumes tumors and their response are heterogeneous and analyzes them at a more local level than current methods.
There is little information to determine treatment efficiency and accuracy during radiotherapy. Currently patients are scanned twice, once up to a week before the start of therapy and a second during or early after therapy. Responses to treatment are then evaluated by comparing pre and post treatment scans through image difference filters proceeded by manual and/or automated rigid image fusion tools. These and other simple comparisons are commonly used but they assume that tumors and their responses are homogenous. These methods are not suitable to fully characterize and detect patterns of molecular modifications that are concealed by low resolution and significant noise levels. Even a visual inspection by an expert does not always identify trends and patterns of tracer activity. A midway approach, where a local rather than a global metric is used to capture fine changes, but a higher level than a voxel-by-voxel analysis is necessary to characterize treatment response in PET images more accurately.
Emory investigators have developed a method to analyze noisy and cluttered datasets using clustering, a technique in which patterns of common traits are extracted by grouping entities with similar characteristics together so that main tendencies or unusual patterns can be identified. This clustering algorithm overcomes acute levels of noise and low signal changes that are characteristic of early changes in tumor response. Compared to classical clustering algorithms, the density-based clustering algorithm defined here allows arbitrary cluster shapes and does not require the input of the number of clusters. The algorithm is especially useful when clusters touch each other because when both cluster centers and cluster boundaries become fuzzy and difficult to determine, the algorithm can clearly identify these centers and boundaries, as well as the degree of membership of each data point within that cluster.
Algorithm has been used on pre- and post-treatment images from patients undergoing radiotherapy.