Human-Guided Design to Explain Deep Learning-based Pneumothorax Classifier

Han Yuan, Peng-Tao Jiang, Gangming Zhao

Keywords: Pneumothorax Diagnosis, Convolutional Neural Network, Explainable Artificial Intelligence, Human-in-the-loop

TL;DR: This paper introduces a human-guided approach to enhance the existing explanation methods for CNNs.

Abstract: Pneumothorax (PTX) is an acute thoracic disease that can be diagnosed with chest radiographs. While deep learning (DL) models have proven effective in identifying PTX on radiographs, they have difficulties in gaining the trust of radiologists if the decision-making logic is unclear. Therefore, various methods have been proposed to explain the PTX diagnostic decision made by DL models. However, several studies indicate that the quality of DL model explanation is suboptimal. This paper introduces a human-guided approach to enhance the existing explanation method. Based on the IoU and Dice between the explanation of model-focusing regions and the ground truth lesion areas, we achieved an increase of 60.6% and 56.5% in Saliency Map, 69.0% and 66.7% in Grad-CAM, and 137.5% and 123.9% in Integrated Gradients.