Local Interpretable Model-Agnostic Explanations for Classification of Lymph Node Metastases.

Marley Maria Bernardes Rebuzzi Vellasco,Eduardo Costa da Silva,Iam Palatnik de Sousa

doi:10.3390/s19132969

Marley Maria Bernardes Rebuzzi Vellasco, Eduardo Costa da Silva + Show 1 more

Open Access

https://doi.org/10.3390/s19132969

Copy DOI

Abstract

An application of explainable artificial intelligence on medical data is presented. There is an increasing demand in machine learning literature for such explainable models in health-related applications. This work aims to generate explanations on how a Convolutional Neural Network (CNN) detects tumor tissue in patches extracted from histology whole slide images. This is achieved using the “locally-interpretable model-agnostic explanations” methodology. Two publicly-available convolutional neural networks trained on the Patch Camelyon Benchmark are analyzed. Three common segmentation algorithms are compared for superpixel generation, and a fourth simpler parameter-free segmentation algorithm is proposed. The main characteristics of the explanations are discussed, as well as the key patterns identified in true positive predictions. The results are compared to medical annotations and literature and suggest that the CNN predictions follow at least some aspects of human expert knowledge.

Highlights

Artificial Intelligence (AI) and Machine Learning (ML) have become increasingly ubiquitous in the study of data from the medical domain
Neural Network (CNN) architectures and other Deep Learning (DL) models for tasks related to medical image classification and segmentation [1,2,3,4]
The Patch Camelyon (P-CAM) dataset has more than 200,000 images, and any of them could be analyzed from the perspective of generating explanations with the above-described procedure

Summary

Introduction

Artificial Intelligence (AI) and Machine Learning (ML) have become increasingly ubiquitous in the study of data from the medical domain. The use of such networks to detect metastases in patches extracted from Whole Slide histopathology Images (WSI) leads to performances that often surpass human pathologists [4,5]. The importance of such results is increased by the fact that pathology experts performing the same task can disagree in more than 20% of the cases [6]. This task is described in the literature as tedious and time consuming, which makes it more prone to human error. The existence of ML/DL systems that can perform the task at very high levels of accuracy seems like a way to overcome this problem

Objectives

Methods

Results

Conclusion