*2.4. Processing Framework*

The processing framework applied to each HS cube is composed by the following steps. First, a standard flat field correction is applied to the images. To this end, the images are transformed from radiance to normalized transmittance by using a reference image that is captured from a blank area of the pathological slide [23]. Then, due to the high correlation of spectral information between adjacent spectral bands, a reduced-band HS image is generated by averaging the neighbors' spectral bands, reducing the number of spectral bands from 826 bands to 275 and slightly reducing the white Gaussian noise. This band reduction is also beneficial for alleviating computational cost in the subsequent image processing. Finally, each image is divided into patches, which will train the CNN. In this section, we will detail the architecture of the proposed neural network, the metrics that are used for performance evaluation, and the proposed data partition scheme.
