Novel Paired Intensity-Pattern Recognition Technique Based On Perfect Intensity-Geometry (PIG)
A Chu*, Q Peng, M Garg, W Tome, Montefiore Medical Center/Albert Einstein college of medicine, Bronx, NY
Presentations
PO-GePV-M-34 (Sunday, 7/25/2021) [Eastern Time (GMT-4)]
Purpose: To present a novel paired intensity-pattern recognition technique using pointwise mutual information (PMI) and image-pair joint probability (PAB) for perfect intensity-geometry (PIG). This technique can facilitate easy pre-/post-treatment comparison for therapy monitoring.
Methods: Mutual information (MI) has been used as an optimization metrics for the intensity-based registration, but it becomes cumbersome during the repetitive classifications. Instead of using MI, PMI-to-PAB map, which describes “the dependence of pair’s intensity pattern” vs. “sub-volume ratio (i.e. PAB)”, was calculated and optimized to approximate to a theoretical reference (PIG). The number of intensity-pattern (Np) of a registered image pair [A, B] was resolved by repetitive classifications (k-means) in the pair’s 2D intensity histogram. The optimal Np was determined in the recognition routine based on the expectation of “PAB/exp(-PMI)” approaching to a unity, e.g. |1-Exp{PAB/exp(-PMI)}|<0.001. Simulated image pairs using a patient’s PTV added with regional intensity variation (to simulate “treatment response”) were tested using this novel approach against functional diffusion map (fDM) method. In addition, the sensitivity of detecting response with the MRI ADC map before and after was also tested.
Results: The novel PIG metrics was robust and was much faster without supervised training sessions compared to the traditional MI (information gain) metrics. It resulted in the patterns of underlying intensity range representing for the same tissues, physio-pathological variants, and mismatches. In the simulation, the PIG method successfully recognized paired intensity-pattern clusters as well as the regional response. In the brain MRI ADC map test, the routine had improved sensitivity to ADC responses compared to optimal fDM (0.844 vs 0.669). Note the traditional fDM can only reach 0.398 with fixed threshold.
Conclusion: Our novel technique using PMI and PAB for perfect intensity-geometry (PIG) offers an unsupervised learning for paired intensity-pattern recognition. This technique can facilitate easy pre- and post-treatment comparison for therapy monitoring.
ePosters
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email