An Explainable Multi-Instance Multi-Label Classification Model for Full Slice Brain CT Images

摘要

Brain CT is the first choice for diagnosing intracranial diseases. However, the doctors who can accurate diagnosis is insufficient with the increasing number of patients. Nowadays, many computer-aided diagnosis algorithms were developed to help doctors diagnose and reduce time. However, most of the research classifies each slice isolated, regard this as an image-level classification problem. It is not comprehensive enough because many conditions can only be diagnosed by considering adjacent slices and the relationships between diseases. In order to better fit the characteristics of this task, we formal it as a Multi-instance Multi-label (MIML) learning problem at sequence-level. In this paper, we analyze the difficulties in the brain CT images classification domain. And we propose an efficient model that can improve performance, reduce the number of parameters and give model explanations. In our model, the convolution neural network (CNN) extracts the feature vector from each image of a set of full slice brain CT. The multi-instance detect module focuses on the key images which could assist doctors quickly locate suspicious images and avoid mistakes. We evaluated our model on two datasets: CQ500 and RSNA. The F1 scores are 0.897 and 0.854 respectively. The proposed model outperforms the previous sequence-level model SDLM with only a quarter of the parameters. Low computation and high performance make the model have clinical applicability.