Table 1.

Detail Steps of proposed UH-DoG.

1. Use a pretrained model to generate a probability map of blobs from original image

2. Initialize the normalization factor γ, and range and step-size of parameter σ, to transform the original image into normalized DoG space.

3. Calculate the Hessian matrix based on normalized DoG smoothed image and generate the Hessian convexity map $HI\left(x,y,z;\sigma \right)$.

4. Calculate average DoG intensity $B_{DoG}(\sigma )=\frac{{\sum }_{(x,y,z)}DoG\left(x,y,z\right)HI(x,y,z;\sigma )}{{\sum }_{(x,y,z)}HI(x,y,z;\sigma )}$ and find the optimum scale section by${\sigma }^{⁎}=argmax{B}_{DoG}\left(\sigma \right)$.

5. Get the optimum Hessian convexity map $HI(x,y,z;{\sigma }^{\ast })$ under scale σ*.

6. Join the probability map with Hessian convexity map to identify true blobs.