Exploiting causality signals in medical images: A pilot study with empirical results

• DOI: 10.1016/j.eswa.2024.123433
• Publication Date: 2024-02-15T12:09:54Z
• Authors: Gianluca Carloni; Sara Colantonio
• Keywords: Feature (linguistics); Feature vector

We present a novel technique to discover and exploit weak causal signals directly from images via neural networks for classification purposes. This way, we model how the presence of feature in one part image affects appearance another different image. Our method consists convolutional network backbone causality-factors extractor module, which computes weights enhance each map according its influence scene. develop architecture variants empirically evaluate all models on two public datasets prostate MRI breast histopathology slides cancer diagnosis. study effectiveness our module both fully-supervised few-shot learning, assess addition existing attention-based solutions, conduct ablation studies, investigate explainability class activation maps. findings show that lightweight block extracts meaningful information improves overall classification, together with producing more robust predictions focus relevant parts That is crucial medical imaging, where accurate reliable classifications are essential effective diagnosis treatment planning.