*ICEM*


8. Form **Ω**(*k*+<sup>1</sup>) = **Ω**(*k*) ∪ ,-- -**B**| (*k*) GFCEM -. Let *k* ← *k* + 1 and go to Step 4.

9. **B**(*k*) CEM is the desired detection abundance fractional map and ICEM is terminated.

Figure 1 delineates how ICEM is processed as a detector where ICEM uses Gaussian filters to smooth CEM-detection abundance fractional maps and feeds back Gaussian-filtered CEM detection abundance fractional maps to provide spatial information for re-processing CEM iteratively. By gradually increasing more spatial information through feedback loops the boundaries of WMH lesions can be detected more effectively. It should also be noted that, if a particular NBE technique is used such as CBEP, then NBE-ICEM can be specified by CBEP-ICEM.

Ω **Figure 1.** A diagram of the *k*th iteration carried out by hyperspectral image classification implementing ICEM on **Ω**(*k*) NBEΩ

#### *2.3. Stopping Rule for ICEM*

To effectively terminate ICEM, DSI defined in [22] as

$$\text{DSI}^{(k)} = \frac{2|S\_k \cap S\_{k-1}|}{|S\_k \cup S\_{k-1}|} \tag{1}$$

is used a stopping criterion where |*S*| is size of a set *S*, *Sk* and *Sk*−<sup>1</sup> are the *k*th thresholded binary image of the *k*th CEM detection abundance fractional map, --**B**|CEM *k* and *k* − 1st thresholded binary image of the *k* − 1st CEM detection abundance fractional map, ---**B**|CEM *k*−1 . Figure 2 depicts a flow chart of a stopping rule using DSI with *ε* as a prescribed error threshold.

**Figure 2.** A flow chart of the stopping rule used for NBE-ICEM.

#### *2.4. Algorithm for NBE-ICEM*

Using Figures 1 and 2, an algorithm developed to implement ICEM in conjunction with NBE can be described as follows. Figure 3 describes a graphic flow chart of implementing NBE-ICEM.
