Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for computer aided diagnosis of a medical image. One of the methods includes processing the medical image through a machine learning (ML) model to provide a first feature representation of the medical image, wherein the ML model includes an input layer and a pooling layer, wherein the first feature representation is an output from the pooling layer, generating, by the ML model, a sequence of second feature representations of the medical image from the first feature representation of the medical image, wherein each second feature representation in the sequence of the second feature representations having a lower dimension than the first feature representation, and generating, by the ML model, an output as a last second feature representation in the sequence of the second feature representations.