Multiple Instance Learning through Explanation by Using a Histopathology Example
Andrei V. Konstantinov, Lev Vladimirovich Utkin · 2022
A new method for solving the multiple instance learning (MIL) problem, which is based on ideas of the black-box model prediction explanation, is proposed. The explanation aims to show instances (pixels, patches) which have the highest contribution into the image (bag) classes and to automatically annotate instances in bags. Three ideas behind the method are used. First, the surrogate black-box model is implemented as the Siamese neural network which is trained on pairs of whole images. Second, patches in each image are changed by using their dynamic fill or noise. Third, noisy images are compared with the original image by using the Siamese neural network such that Euclidean distances between outputs of the network depending on the noise level form a shape function for every patch. The shape function is interpreted from its contribution into the image class. Numerical experiments with the real Breast Cancer Cell Segmentation dataset illustrate the method.