Attention-based full slice brain CT image diagnosis with explanations

Abstract

Diagnose brain diseases by brain CT images is one of the most common ways. But, it usually takes several (>7) years to train a professional doctor because it is very challenging to diagnose brain diseases correctly. The study on automated assistance of brain CT diagnosis is still limited. In this paper, we research the challenges of this task and propose a method by simulating human doctor diagnosis habits. Our method analyzes a full slice of brain CT images, instead of every single one, to take into account continuous changes of the whole brain structure, simulate the way the doctor diagnoses. To avoid redundancies in a thin slice scan, we propose a redundancy removal and data augmentation method that can both reduce computation complexity and improve performance without information loss. Doctors make a diagnosis by observing several key images and key points in them. Our method achieved this by two steps of attention mechanisms. It can highlight the images and key points that have significant impacts on the prediction and explain the results. We evaluated our method on two public datasets CQ500 and RSNA, which achieved 0.9262 and 0.8650 F1 score respectively. Moreover, an experienced doctor (with 29 years of experience) verified the promising clinical application value of the proposed method through manual experiments.