Similarity/Dissimilarity of Image (Signal) Comparison with Changing Random Variable Distributions
Presentations
PO-GePV-M-343 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: To introduce an information gain (IG) metrics that can adapt changing random variable pair (RVP) distributions for Image comparison.
Methods: M1: A mutual information (MI) in Kullback-Leibler divergence (KLD) framework evaluates the “dissimilarity” of the P(X,Y) from P(X)P(Y). We define “baseline” as P(X,Y)=P(X)P(Y). For the case of MI, the baseline MI is lower bound (LB) – an independent RVP. M2: We propose to replace the baseline of MI with MI’s upper bound (UB), P(X,Y)=exp[-PMI(X,Y)], where PMI stands for pointwise mutual information. The baseline is parameterized by P(X,Y) and PMI without RVP distributions. Experiments:(i)To test “M1” and “M2” for searching an optimal decorrelation level for heart rate (HR), and also compare against with the correlation coefficient (CC) under the association of a sleeping stage (SS) by ECG/EOG as the ground truth with an optimally decorrelated HR. (ii)To compare the recognizable abnormalities by “M2” with optimal-threshold (Op_Theshold) discriminator. The ground truth is the “abnormalities” added to a GBM patient’s brain ADC image with a gaussian intensity shift mean=342x10-6, sigma=375 x10-6 (mm2/sec) to 60% of randomly selected voxels within the patient’s white matter (WM).
Results: (1) M2 matches the CC for predicting SS by the best decorrelated HR. M1 cannot match due to one RVP’s distributions changing in the decorrelation process.(2) Detectability for abnormalities for Op_Theshold/ M2. Accuracy: 0.891/0.867, Precision: 0.767/0.646, Sensitivity: 0.670/0.844, Specificity: 0.948/0.874. Note: “Op_Threshold” represents the best threshold-discriminator of predictability if a ground-truth is provided for detection. The detectability of M2 is comparable to the best threshold-discriminator. Particularly, sensitivity of M2 is much better than that of Op_Theshold.
Conclusion: Result(1): the success of M2 is due to its non-zeroed baseline that is parameterizable without marginal probabilities. Result(2):The perfect correlation character of UB-baseline offers perfect intensity-geometry paired recognition.
Keywords
Image Correlation, Linear Discriminant Analysis, Mutual Information
Taxonomy
IM/TH- Informatics: Informatics in Imaging (general)
Contact Email