Medical Image Retrieval with Relevance Feedback via Pairwise Constraint Propagation

( Menglin Wu ),( Qiang Chen ),( Quansen Sun ) 한국인터넷정보학회 2014 KSII Transactions on Internet and Information Syst Vol.8 No.1

Relevance feedback is an effective tool to bridge the gap between superficial image contents and medically-relevant sense in content-based medical image retrieval. In this paper, we propose an interactive medical image search framework based on pairwise constraint propagation. The basic idea is to obtain pairwise constraints from user feedback and propagate them to the entire image set to reconstruct the similarity matrix, and then rank medical images on this new manifold. In contrast to most of the algorithms that only concern manifold structure, the proposed method integrates pairwise constraint information in a feedback procedure and resolves the small sample size and the asymmetrical training typically in relevance feedback. We also introduce a long-term feedback strategy for our retrieval tasks. Experiments on two medical image datasets indicate the proposed approach can significantly improve the performance of medical image retrieval. The experiments also indicate that the proposed approach outperforms previous relevance feedback models.

설명가능한 인공지능 기술을 이용한 인공신경망 기반 수질예측 모델의 성능향상

이원진,이의훈 한국수자원학회 2023 한국수자원학회논문집 Vol.56 No.11

최근 인공신경망(Artificial Neural Network, ANN)의 연구가 활발하게 진행되면서 ANN을 이용하여 하천의 수질을 예측하는 연구가 진행되고 있다. 그러나 ANN은 Black-box의 형태이기 때문에 ANN 내부의 연산과정을 분석하는데 어려움이 있다. ANN의 연산과정을 분석하기 위해 설명가능한 인공지능(eXplainable Artificial Intelligence, XAI) 기술이 사용되고 있으나, 수자원 분야에서 XAI 기술을 활용한 연구는 미비한 실정이다. 본 연구는 XAI 기술 중 Layer-wise Relevance Propagation (LRP)을 사용하여 낙동강의 다산 수질관측소의 수온, 용존산소량, 수소이온농도 및 엽록소-a를 예측하기 위한 Multi Layer Perceptron (MLP)을 분석하였다. LRP를 기반으로 수질을 학습한 MLP를 분석하여 수질을 예측하기 위한 최적의 입력자료를 선정하고, 최적의 입력자료를 이용하여 학습한 MLP의 예측결과에 대한 분석을 실시하였다. LRP를 이용하여 최적의 입력자료를 선정한 결과를 보면, 수온, 용존산소량, 수소이온농도 및 엽록소-a 모두 주변지역의 일 강수량을 제외한 입력자료를 학습한 MLP의 예측정확도가 가장 높았다. MLP의 용존산소량 예측결과에 대한 분석결과를 보면, 최고점에서 수소이온농도 및 용존산소량의 영향이 크고 최저점에서는 수온의 영향이 큰 것으로 분석되었다. Recently, as studies about Artificial Neural Network (ANN) are actively progressing, studies for predicting water quality of rivers using ANN are being conducted. However, it is difficult to analyze the operation process inside ANN, because ANN is form of Black-box. Although eXplainable Artificial Intelligence (XAI) is used to analyze the computational process of ANN, research using XAI technology in the field of water resources is insufficient. This study analyzed Multi Layer Perceptron (MLP) to predict Water Temperature (WT), Dissolved Oxygen (DO), hydrogen ion concentration (pH) and Chlorophyll-a (Chl-a) at the Dasan water quality observatory in the Nakdong river using Layer-wise Relevance Propagation (LRP) among XAI technologies. The MLP that learned water quality was analyzed using LRP to select the optimal input data to predict water quality, and the prediction results of the MLP learned using the optimal input data were analyzed. As a result of selecting the optimal input data using LRP, the prediction accuracy of MLP, which learned the input data except daily precipitation in the surrounding area, was the highest. Looking at the analysis of MLP's DO prediction results, it was analyzed that the pH and DO a had large influence at the highest point, and the effect of WT was large at the lowest point.

Multimodal Perturbation and Cluster Pruning Based Selective Ensemble Classifier and Its Iron Industrial Application

Qiannan Wu,Yifei Sun,Xuefeng Yan,Lihua Lv 제어·로봇·시스템학회 2023 International Journal of Control, Automation, and Vol.21 No.11

The selective ensemble aims to search the optimal subset balanced accuracy and diversity from the original base classifier set to construct an ensemble classifier with strong generalization performance. A selective ensemble classifier named BRFS-APCSC is proposed in this paper, which realizes the generation and selection of a set of accurate and diverse base classifiers respectively. In the first step, a multimodal perturbation method is introduced to train distinct base classifiers. The method perturbs the sample space by Bootstrap and disturbs the feature space under a newly proposed semi-random feature selection, which is a combination of the core attribute theory and the improved maximum relevance minimum redundancy algorithm. Then, to search the optimal classifier subset, affinity propagation clustering is added to cluster base classifiers in the first step, then the base classifiers are regarded as features so that the improved maximum relevance minimum redundancy algorithm is applied to select parts of base classifiers from each cluster for integration. UCI datasets and an actual dataset of semi-decarbonization are employed to verify the performance of BRFS-APCSC. The experimental results demonstrate that BRFS-APCSC has significantly difference with other selective ensemble methods and improve the classification accuracy.

DNN과 계층 연관성 전파를 이용한 PM2.5 고농도 사례의 인자 중요도 분석

유숙현 한국멀티미디어학회 2023 멀티미디어학회논문지 Vol.26 No.8

In this study, we used Layer-wise Relevance Propagation (LRP) to analyze the level of contribution of input factors to the predictive results of the PM2.5 predictive model. First, we trained the DNN prediction model using data from 2015 to 2020, and then evaluated it using data from 2021. Next, we performed LRP on the evaluation data to analyze the importance of input factors in the prediction results. As a result, factors with consistently high importance regardless of concentration were O_TA, O_TD, O_RH, O_U, O_V, and O_PA, whereas PM10 and O_RN_ACC were observed to have lower importance. Furthermore, to analyze the characteristics of high-concentration data that are generally difficult to predict compared to low-concentration data, we divided the data by concentration and analyzed the importance of input factors. As a result, the importance of O_PM2.5 was high in the high concentration pattern and the importance of O_radiation was low, while the opposite trend was observed in the low concentration pattern. In particular, for high-concentration patterns that started suddenly and lasted more than three days, we analyzed the importance of input factors by time and factor. These high-concentration patterns with these characteristics showed significantly increased importance in the O_PM2.5 factor in the T12 interval closest to the prediction time, and it was observed that the importance of the F_PM2.5 factor increased slightly. Applying the factor importance results analyzed in this study to the PM2.5 prediction model is expected to improve prediction accuracy for high concentration patterns that are difficult to predict compared to general patterns.

Untangling the contribution of input parameters to an artificial intelligence PM<SUB>2.5</SUB> forecast model using the layer-wise relevance propagation method

Dasol Kim,Chang-Hoi Ho,Ingyu Park,Jinwon Kim,Lim-Seok Chang,Min-Hyeok Choi 한국기상학회 2022 한국기상학회 학술대회 논문집 Vol.2022 No.10

XAI(eXplainable AI) 기반 하수처리 활성오니공정 용존산소 제어를 위한 신경회로망 모델링

남의석 대한전기학회 2022 전기학회논문지 Vol.71 No.8

In this paper, we proposed Dissolved Oxygen(DO) neural network model of activated sludge process using XAI(eXplainable AI) in wastewater treatment system. To improve the model performance, input water qualities are to be reliable and have a much influences in DO biological operation. In regulations, COD, T-N, T-P, pH, SS of effluent are hourly to transmitted in Korea Environment Corporation. If these values are exceed the legal standards, the penalty is given. Therefore these data are very reliable and is monitored by operators critically. So these data is to be inputs of DO neural network model. And XAI(eXplainable AI) is utilized to decide which input water qualities have much influences in the process. LRP(Layer-wise Relevance Propagation) is used among various XAI(eXplainable AI) methods. NH4, MLSS, pH in aeration tank and COD, TN, TP, SS in secondary clarifier are input candidates of model for Do neural network modeling. Using LRP, COD, NH4, MLSS, SS are decided to be inputs of Do neural network model. The validity of the proposed method was proved by applying to the DO neural network model of activated sludge process which was developed in previous research. 3 years hourly data was used for modeling and estimation. The result show that the performance of the proposed model was improved in comparison of conventional neural network models. In the future, absolute values of weight in LRP will be more considered because we considered only the inputs orders of influencing on DO biological operation.

계층 관련성 전파와 DNN을 이용한 PM<SUB>2.5</SUB> 예보 지수별 입력 인자의 기여도 특성 분석

유숙현(SukHyun Yu) 한국멀티미디어학회 2024 멀티미디어학회논문지 Vol.27 No.5

In this study, we developed a DNN forecast model to predict PM<SUB>2.5</SUB> for the next three days and used Layer-wise Relevance Propagation to investigate the contributions of input factors to the forecast results. We classified the prediction results into 16 indices to assess the contributions of input factors. As a result, pressure, temperature, dew point, relative humidity, U, and V showed the highest contributions across indices, regardless of the specific index. Followed by O₂, NO₂, and CO exhibited high contributions, while accumulated precipitation and PM10 had the lowest contributions. Radiation and O₃ showed high contributions in low-concentration indices, but conversely exhibited low contributions in high-concentration indices. Low-concentration index ranges showed similar contributions to the overall index, while in high-concentration index ranges, the contribution of forecast PM<SUB>2.5</SUB> at the forecast time slot increased. Index ranges where low concentrations were incorrectly forecasted resembled high-concentration indices except for relative humidity, NO₂, and CO, whereas index ranges where high concentrations were incorrectly predicted exhibited significant temporal variations in the contribution of radiation. These characteristics of input factor contributions to forecast results are expected to be utilized for data quality improvement and forecast performance enhancement.

Estimating PM2.5 concentration of the conterminous United States via interpretable convolutional neural networks

Park, Yongbee,Kwon, Byungjoon,Heo, Juyeon,Hu, Xuefei,Liu, Yang,Moon, Taesup Elsevier 2020 Environmental pollution Vol.256 No.-

<P><B>Abstract</B></P> <P>We apply convolutional neural network (CNN) model for estimating daily 24-h averaged ground-level P <SUB> M 2.5 </SUB> of the conterminous United States in 2011 by incorporating aerosol optical depth (AOD) data, meteorological fields, and land-use data. Unlike some of the recent supervised learning-based approaches, which only utilized the predictors from the location of which P <SUB> M 2.5 </SUB> value is estimated, we naturally aggregate predictors from nearby locations such that the spatial correlation among the predictors can be exploited. We carefully evaluate the performance of our method via overall, temporally-separated, and spatially-separated cross-validations (CV) and show that our CNN achieves competitive estimation accuracy compared to the recently developed baselines. Furthermore, we develop a novel predictor importance metric for our CNN based on the recent neural network interpretation method, Layerwise Relevance Propagation (LRP), and identify several informative predictors for P <SUB> M 2.5 </SUB> estimation.</P> <P><B>Highlights</B></P> <P> <UL> <LI> Convolutional neural network (CNN) accurately estimates daily averaged PM2.5. </LI> <LI> Layerwise relevance propagation (LRP) is used to obtain predictor impor-tance list. </LI> <LI> Exploiting spatial correlation of nearby predictors boosts the estimation accuracy. </LI> <LI> Weighted average feature of PM2.5 is useful even when CNN is used. </LI> <LI> CNN can generate smooth annual prediction map of PM2.5 for the con-terminous US. </LI> </UL> </P> <P><B>Graphical abstract</B></P> <P>[DISPLAY OMISSION]</P>

Methods for interpreting and understanding deep neural networks

Montavon, Gré,goire,Samek, Wojciech,Mü,ller, Klaus-Robert Elsevier 2018 Digital signal processing Vol.73 No.-

<P><B>Abstract</B></P> <P>This paper provides an entry point to the problem of interpreting a deep neural network model and explaining its predictions. It is based on a tutorial given at ICASSP 2017. As a tutorial paper, the set of methods covered here is not exhaustive, but sufficiently representative to discuss a number of questions in interpretability, technical challenges, and possible applications. The second part of the tutorial focuses on the recently proposed layer-wise relevance propagation (LRP) technique, for which we provide theory, recommendations, and tricks, to make most efficient use of it on real data.</P>

C-rank: 웹 페이지 랭킹을 위한 기여도 기반 접근법

이상철(Sang-Chul Lee),김동진(Dong-Jin Kim),손호용(Ho-Yong Son),김상욱(Sang-Wook Kim),이재범(Jae Bum Lee) 한국정보과학회 2010 정보과학회 컴퓨팅의 실제 논문지 Vol.16 No.1

수많은 웹 문서로부터 웹 서퍼가 원하는 정보를 찾기 위해 다양한 검색 엔진들이 개발되어왔다. 검색 엔진에서 가장 중요한 기능 중 하나는 사용자 질의에 대해서 웹 문서를 평가하고 랭킹을 부여하는 것이다. PageRank등의 기존 하이퍼링크 정보를 이용한 웹 랭킹 알고리즘은 토픽 드리프트 현상을 발생시킨다. 이러한 문제를 해결하기 위하여 연관성 파급 모델이 제안되었지만, 기존의 연관성 파급 모델을 기반으로 하는 랭킹 알고리즘은 성능상의 이유로 실제 웹 검색 엔진에서 사용하기 어렵다. 본 논문에서는 이러한 토픽 드리프트 현상을 완화하면서 좋은 성능을 제공하는 새로운 랭킹 알고리즘을 제안한다. 다양한 실험을 통하여 기존 알고리즘들과 비교한 제안하는 알고리즘의 우수성을 검증한다. In the past decade, various search engines have been developed to retrieve web pages that web surfers want to find from world wide web. In search engines, one of the most important functions is to evaluate and rank web pages for a given web surfer query. The prior algorithms using hyperlink information like PageRank incur the problem of ‘topic drift’. To solve the problem, relevance propagation models have been proposed. However, these models suffer from serious performance degradation, and thus cannot be employed in real search engines. In this paper, we propose a new ranking algorithm that alleviates the topic drift problem and also provides efficient performance. Through a variety of experiments, we verify the superiority of the proposed algorithm over prior ones.