The natural language explanation algorithms for the lung cancer computer-aided diagnosis system

• Published in: Artificial intelligence in medicine Vol. 108; p. 101952
• Main Authors: Meldo, Anna, Utkin, Lev, Kovalev, Maxim, Kasimov, Ernest
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.08.2020
• Subjects: Causability; Classification; Computer-aided diagnosis; Explainable AI; LIME; Lung cancer; Natural language; Visual explanations; Causability; Explainable AI; LIME; Lung cancer; Classification; Natural language; Visual explanations; Computer-aided diagnosis
• ISSN: 0933-3657; 1873-2860; 1873-2860
• DOI: 10.1016/j.artmed.2020.101952

•Two algorithms of the lung cancer CADx system explanation is proposed.•The explanation is implemented via natural language.•A low-dimensional feature representation of CT lung nodule images is considered to simplify classification and explanation.•An efficient explanation method is based on using a set of simple classifiers which provide short phrases as classes. Two algorithms for explaining decisions of a lung cancer computer-aided diagnosis system are proposed. Their main peculiarity is that they produce explanations of diseases in the form of special sentences via natural language. The algorithms consist of two parts. The first part is a standard local post-hoc explanation model, for example, the well-known LIME, which is used for selecting important features from a special feature representation of the segmented lung suspicious objects. This part is identical for both algorithms. The second part is a model which aims to connect selected important features and to transform them to explanation sentences in natural language. This part is implemented differently for both algorithms. The training phase of the first algorithm uses a special vocabulary of simple phrases which produce sentences and their embeddings. The second algorithm significantly simplifies some parts of the first algorithm and reduces the explanation problem to a set of simple classifiers. The basic idea behind the improvement is to represent every simple phrase from vocabulary as a class of the “sparse” histograms. An implementation of the second algorithm is shown in detail.