Clustered-CAM: Visual Explanations for Deep Convolutional Networks for Thyroid Nodule Ultrasound Image Classification

Ali Eskandari, Hongbo Du, ALAA ALZOUBI

Keywords: Deep Learning Visualization, Thyroid Cancer Recognition, Ultrasonography.

TL;DR: This study presents a new method (Clustered-CAM) to investigate the efficacy of applying ablation on a group of similar feature maps for accurate saliency maps generation.

Abstract: Explaining the CNN classification decision is crucial for the system acceptance in critical applications such as tumour recognition in 2D Ultrasound images. Generating saliency maps that highlight the image regions contributing to the final CNN decision is one of the most common techniques. In this paper, we propose a clustering-based approach to group similar feature maps before assigning importance scores to produce a more accurate and less sensitive visual explanation for CNN models for thyroid nodule classification in US images. Our study with a dataset of 864 ultrasound images shows that the Clustered-CAM achieved a lower average drop and higher percent increase in confidence comparing to the-state-of-the-art techniques. We demonstrate that Clustered-CAM is an effective and promising approach for visualising the CNN model decisions for thyroid nodule recognition.

Registration: I acknowledge that acceptance of this work at MIDL requires at least one of the authors to register and present the work during the conference.

Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.

Paper Type: novel methodological ideas without extensive validation

Primary Subject Area: Interpretability and Explainable AI

Secondary Subject Area: Application: Radiology

Confidentiality And Author Instructions: I read the call for papers and author instructions. I acknowledge that exceeding the page limit and/or altering the latex template can result in desk rejection.