A H UMAN -CENTERED EXPLAINABLE AI F RAMEWORK FOR BURN DEPTH CHARACTERIZATION A P REPRINT

2025-04-30 0 0 2.02MB 8 页 10玖币

侵权投诉

A HUMAN-CENTERED EXPLAINABLE AI FRAMEWORK FOR

BURN DEPTH CHARACTERIZATION

A PREPRINT

Maxwell J. Jacobson

Purdue University

jacobs57@purdue.edu

Daniela Chanci Arrubla

Emory University

daniela.chanci.arrubla@emory.edu

Maria Romeo Tricas

Purdue University

mromeotr@purdue.edu

Gayle Gordillo

Indiana University

gmgordil@iu.edu

Yexiang Xue

Purdue University

yexiang@purdue.edu

Chandan Sen

Indiana University

cksen@iu.edu

Juan Wachs

Purdue University

jpwachs@purdue.edu

January 3, 2023

ABSTRACT

Approximately 1.25 million people in the United States are treated each year for burn injuries. Precise

burn injury classiﬁcation is an important aspect of the medical AI ﬁeld. In this work, we propose an

explainable human-in-the-loop framework for improving burn ultrasound classiﬁcation models. Our

framework leverages an explanation system based on the LIME classiﬁcation explainer to corroborate

and integrate a burn expert’s knowledge — suggesting new features and ensuring the validity of the

model. Using this framework, we discover that B-mode ultrasound classiﬁers can be enhanced by

supplying textural features. More speciﬁcally, we conﬁrm that texture features based on the Gray

Level Co-occurance Matrix (GLCM) of ultrasound frames can increase the accuracy of transfer

learned burn depth classiﬁers. We test our hypothesis on real data from porcine subjects. We show

improvements in the accuracy of burn depth classiﬁcation — from 88% to 94% — once modiﬁed

according to our framework.

Keywords

Burn Analysis

Ultrasound

Deep Learning

Computer Vision

Explainability

Human-in-the-loop

1 Introduction

Our work focuses on the task of burn depth estimation. Performing this task accurately can be critical for the welfare of

burn victims, but it is also a challenge due to the high variability in the visual appearance of burns. Different imaging

modalities have been explored to solve this problem and improve the accuracy of the diagnosis Sen et al. [2016] —

e.g. color photographs, ultrasound, infrared thermography, laser speckle imaging, and laser doppler imaging Thatcher

et al. [2016]. Recent studies combine imaging modalities with machine learning and deep learning models. Cirillo et

al Cirillo et al. [2019] compared the performance of VGG16 Simonyan and Zisserman [2015], GoogleNet Szegedy

et al. [2015], ResNet50 He et al. [2016], and ResNet101, pretrained on ImageNet Deng et al. [2009], to classify their

labeled dataset of RGB burn images. Similarly, Chauhan & Goyal Chauhan and Goyal [2020] carried out the burn

depth classiﬁcation based on the speciﬁc characteristics of the body region in which the injury was located. However,

RGB images can negatively inﬂuence the accuracy of predictions due to lighting conditions, skin color, or general

variability in wound presentation. Besides, RGB images, optical coherence tomography Singla et al. [2018], spatial

frequency-domain imaging Rowland et al. [2019], and ultrasound Lee et al. [2020] have been utilized for feature

extraction. Harmonic B-mode ultrasound (HUSD) — a non-invasive sound-based imaging technique — is used in

this work. B-mode ultrasound is based on the transmission of small pulses of ultrasound. Echoes reﬂected back to

the transducer from body tissues that have different acoustic impedances, which can be measured to build a 3D map

of tissues. Second harmonic frequency echoes are used in order to reduce the artifacts in the image produced by the

reﬂection of echoes at different frequencies Narouze [2018].

arXiv:2210.13535v2 [cs.CV] 2 Jan 2023

Human-centered XAI for Burn Depth Characterization A PREPRINT

Figure 1: Our human-in-the-loop through explainability framework. A human possessing expert knowledge trains a

classiﬁer. Our LIME-based explainer provides insights to the human based on the current model. The human applies

expert knowledge to interpret the explanation. The human expert modiﬁes the model appropriately, and the process can

be repeated. In our case, we modify the network by adding new features based on expert medical knowledge regarding

burns.

Computer vision based techniques lack the human expertise in current medical ML models. Therefore, a human-in-the-

loop system built on Explainable Artiﬁcial Intelligence is proposed here. Human-in-the-loop (HITL) is an Artiﬁcial

Intelligence (AI) paradigm that assumes the presence of human experts that can guide the learning or operation of

the otherwise-autonomous system. Lundberg et al. Lundberg et al. [2018] developed and tested a system to prevent

hypoxaemia during surgery by providing anaesthesiologists with interpretable hypoxaemia risks and contributing

factors. Later, Sayres et al. Sayres et al. [2019] proposed and evaluated a system to assist diabetic retinopathy grading

by ophtalmologists using a deep learning model and integrated gradients explanation.

An Explainable Artiﬁcial Intelligence (XAI) is an intelligent system which can be explained and understood by a human

Gohel et al. [2021]. For this, we utilize LIME (Local Interpretable Model-agnostic Explanations) Ribeiro et al. [2016],

a recent method that is able to explain the predictions of any classiﬁer model in an interpretable manner. LIME operates

by roughly segmenting an image into feature regions, then assigning saliency scores to each region. Higher scoring

zones are more important in arriving at the classiﬁcation result of the studied model. The algorithm ﬁrst creates random

permutations of the image to be explained. Then, the classiﬁcation model is run on those samples. Distances between

those samples and the original image can be calculated, which are then converted to weights by being mapped between

zero and one using a kernel function. Finally, a simple linear model is ﬁtted around the initial prediction to generate

explanations. This explanation provided by LIME is the result of the following minimization:

ξ(x) = argmin

g∈G

L(f, g, πx) + Ω(g)(1)

Let

and

be the image and classiﬁer to be examined and

as the class of interpretable models like decision trees and

linear models. The complexity of the model (e.g. depth of a decision tree, number of non-zero weights in a linear model)

should be as small as possible to maintain explainability. This complexity can be deﬁned as

Ω(g)

. Explanations by

LIME are found by ﬁtting explainable models — minimizing the sum of the local faithfulness loss

and the complexity

score. Permutated sampling is used to approximate this local faithfulness loss. For this reason, a proximity measure

πx(z)

calculates the distance between

and another image

. The objective is minimizing the ﬁdelity function while

maintaining a measure of complexity low enough to be interpretable. This minimization makes no assumptions about

in order to generate a model-agnostic explanation.

Explainable intelligence is useful when combined with HITL systems because they provide understandable and

qualitative information about the relationship between the instance’s components and the model’s prediction. Therefore,

an expert can make an informed decision about whether the model is reliable, and can make the necessary changes if it

is not — eventually reaching a conﬁdent result. This is extremely important, above all in the medical ﬁeld, because of

the severe ethical implications that suppose a wrong medical diagnosis.

By deploying our explainable human-in-the-loop method, we were able to conﬁrm the importance of one family of

features which can enhance a convolutional burn prediction classiﬁer — statistical texture. From the Gray Level

Co-ocurrence Matrix (GLCM) — a method that represents the second-order statistical information of gray levels

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

AHUMAN-CENTEREDEXPLAINABLEAIFRAMEWORKFORBURNDEPTHCHARACTERIZATIONAPREPRINTMaxwellJ.JacobsonPurdueUniversityjacobs57@purdue.eduDanielaChanciArrublaEmoryUniversitydaniela.chanci.arrubla@emory.eduMariaRomeoTricasPurdueUniversitymromeotr@purdue.eduGayleGordilloIndianaUniversitygmgordil@iu.eduYexiangXueP...

展开>> 收起<<

A H UMAN -CENTERED EXPLAINABLE AI F RAMEWORK FOR BURN DEPTH CHARACTERIZATION A P REPRINT.pdf

共8页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

A H UMAN -CENTERED EXPLAINABLE AI F RAMEWORK FOR BURN DEPTH CHARACTERIZATION A P REPRINT

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: