Wavelet-Based Prototype Learning for Medical Image Classification

The need for transparent decisions in safety-critical domains is a burning necessity, specifically given the widespread adoption of intelligent systems in diverse areas, like e.g, medical data processing. In this paper, we continue to explore the prototype-based networks, which provide interpretable predictions of the input image. The mechanism is via prototypes, a structural part of the decision-making system. We apply the interpretable architecture - named WaveProtoPNet - on medical image data and explore the prototype-based architecture. We show the interpretability and the reduced parameter size of the model, but also focus on its performance on the NCT-CRC-HE-IOOK dataset on human tissue types. By using wavelets, the parameters are reduced substantially, leading to significant saving in both memory and computation. As with all prototype-based methods, interpretability is a consequence of prototype locations, providing an “explanatory” insight into the classification. Our model remains highly adaptable, as images with custom sizes could be learnt.