Towards Translational Incremental Similarity-Based Reasoning in

the Pathology Department of Singapore National University Hospital, for the .... In our opinion, CBIR of today doesn't only organize digital data, as it was the ...
Télécharger le PDF
57KB taille 3 téléchargements 349 vues
commentaire
Report
Towards Translational Incremental Similarity-Based Reasoning in Breast Cancer Grading Adina Eunice Tutac1,4, Daniel Racoceanu1,5, Wee-Kheng Leow1,3, Henning Mülller6,7 Thomas Putti2,Vladimir Cretu4 1

Image Perception, Access & Language IPAL (UMI CNRS 2955,UJF,NUS,I2R) Singapore, 2 National University Hospital, Singapore, 3National University of Singapore, 4 Politehnica University of Timisoara Romania, 5University of Besançon France, 6 Medical Informatics Service, University Hospitals of Geneva, Switzerland 7 University of Applied Sciences Western Switzerland, Sierre, Switzerland [email protected], [email protected], [email protected], [email protected], [email protected], [email protected]

ABSTRACT One of the fundamental issues in bridging the gap between the proliferation of Content-Based Image Retrieval (CBIR) systems in the scientific literature and the deficiency of their usage in medical community is based on the characteristic of CBIR to access information based on single images, only. Yet, the way physicians are reasoning about patients leads intuitively to a case representation, this implying information structured by cases. Hence, a proper solution to overcome this gap is to consider an approach similar to Case-Based Reasoning (CBR), which naturally introduces medical knowledge structured by cases. Moreover, the knowledge is incrementally learned in a CBR system. Apart from these CBR advantages, CBIR is currently focused on image level, only. Therefore, we advance the idea of a translational incremental similarity-based reasoning (TISBR) approach. The purpose of this study is to find a translational solution from theoretical ideas to clinical practice. The reasoning is based on combined characteristics of CBIR and CBR: incremental learning of medical knowledge, structured by cases (CBR), the usage of image in similar cases (CBIR), concept of similarity (central for both paradigms). For this purpose, three major axes are explored: the indexing, cases retrieval and the search refinement for Breast Cancer Grading (BCG), a powerful prognosis indicator worldwide. The effectiveness of this strategy is currently evaluated over cases provided by the Pathology Department of Singapore National University Hospital, for the indexing axis. With an accuracy of currently 80%, TISBR launches interesting perspectives for complex reasoning in medical research, mapped to specific applications. Description of purpose Apart from Content-Based Image Retrieval’s and Case- Based Reasoning’s individual promising applicability in medical communities, both face significant issues with respect to clinical practice3, 6, and inherent questions are raising above. Is it efficient enough to analyze only images, for identifying relevant features in a retrieval process? Is the CBIR process the end of a medical procedure in order to give a good diagnosis or prognosis? On the other hand, is textual description of a case representation accurate enough to provide a good diagnosis as a valid solution? Would a combined approach solve their limitations? Therefore, to answer these questions, our paper presents a comparative analysis of both CBIR and CBR with the scope of identifying common characteristics and advantages that can be used to propose an integrated framework: a translational incremental similarity-based reasoning (TISBR) for Breast Cancer Grading. Textual information describing symptoms and prognosis - as it is a commonly representation of a case in CBR - combined with image content analysis- typical for CBIR , as well as the concept of similarity, common for both , give us the insight for the hybrid approach. Methods 1. The indexing A general comparison emphasizing principles and techniques is illustrated by Table1. We developed a feature-based method for a semi-automated knowledge-guided semantic indexing process, thus combining characteristics of both CBIR and CBR. The key idea is to use the medical knowledge (MK-tubule formation, nuclear pleomorphism and mitosis count criteria for grading) in terms of concepts and rules, for creating a Computer Vision

(CV) concepts and rules correspondence8. For this aim, we structure the knowledge (specific to CBR) using OWL-DL sublanguage, in a BCG ontology model validated under Pellet reasoner. Indexing Indexing Principles

Content-Based Image Retrieval1, 2, 7  purpose oriented

Similar Indexing Techniques

   

Different Indexing Techniques Characteristics

feature-based structural features salient-features learning-based

feature indexing

Case-Based Reasoning10, 11, 14  predictive  purpose oriented  abstract/concrete enough  features &dimensions  difference-based  inductive learning  similarity & explanation-based case indexing

Table1. Indexing in CBIR & CBR

The structured information is then used for the image processing and analysis step, (specific to CBIR). The features vector, similar with a case indexing in terms of CBR, contains the features-symptoms (TubuleFormationROI, MitosisROI, NucleiROI) and their values, together with the prognosis, defined as the local and global grading (per frame/per entire histopathology slide). 2. The second axis is the retrieval. Similarly, a generic analysis of retrieval in CBIR and CBR is given by Table2. Retrieval Content-Based Image Retrieval3, 5 Case-Based Reasoning10,11, 14 Retrieval Principles Similar Retrieval Techniques Different Retrieval Techniques

 criteria selection     

Query-By-Semantic-Example semantic retrieval Nearest-neighbor retrieval Query-by-Keyword Query-by-Visual-Example(QBVE)

 criteria selection  memory model  knowledge-guided  Nearest-neighbor retrieval  inductive  validated retrieval

Table2. Retrieval in CBIR and CBR

3. The refinement in CBIR and CBR is presented by Table3. In CBIR is identified as the relevance feedback step, whereas in CBR the case adaptation plays the role of refinement. The difference between the two approaches appears with respect to the target of refinement: in CBIR, the query is to be refined, whilst in CBR, the solution is refined. Relevance feedback/ Case adaptation Refinement Principles

Content-Based Image Retrieval2, 7

Similar Refinement Techniques

 no RF/naïve RF

Different Refinement Techniques Characteristics

 structural

 feature re-weighting  specialized user-driven  memory-retrieval  active-learning  probabilistic query refinement

Case-Based Reasoning10, 11  structural  derivational  null adaptation  parameter adjustment  critic-based  model-based  abstraction & respecialization/reinstantiation  derivational replay  case base substitution solution refinement

Table3. Relevance feedback/Case adaptation in CBIR & CBR

Novelty 1. Content- Based Image Retrieval & Case- Based Reasoning as methodologies

Although, both Content-Based Image Retrieval and Case-Based Reasoning are conceptually referred as technologies1, 2, 9, we adopt a similar approach as Watson13, considering CBR organized as a set of principles which guide techniques to be used in problem solving rather than an isolated technique specialized for only very specific tasks12. Moreover, we propel CBIR at the same level, of methodology. Our point of view is depicted by Figure1.

Figure1. CBIR & CBR. Methodology versus Technology

Methodology versus Technology Content- Based Image Retrieval

Definition

Reason

Implication

Technology 2,7

set of methods to solve problems1

 manifold applications

Methodology

continuous development

Technology 9

AI Technology description 9

Methodology 13

set of principles to solve problems12

 knowledge-based systems flexibility  task limitation  research limitation  can use any technology  hybrid systems  future research

Case-Based Reasoning

Table4. Reasons and Implications for Methodology/Technology in CBIR & CBR

We envisage CBIR made significant development in the recent years in terms of concepts, techniques and application domain. In our opinion, CBIR of today doesn’t only organize digital data, as it was the main objective in its early years due to the semantic web development. From this standpoint, we consider that the principles of CBR can also be applied to CBIR. Table4 highlights the reasons and implications of CBIR and CBR technology versus methodology. 2. Global analysis of CBIR and CBR from indexing, retrieval and refinement perspectives for TISBR approach Translational Incremental Similarity-based Reasoning (TISBR) proposes a general analysis of Content-Based Image Retrieval and Case-Based Reasoning from three perspectives: indexing, retrieval and refinement. The purpose is to find a solution to translate theoretical ideas into clinical practice for the BCG application. We do that by identifying CBIR’s and CBR’s common characteristics (both are similarity-based) and advantages: image processing and analysis (CBIR), incrementally learning of medical knowledge (CBR) to develop a fusion approach. 3. Ontologies for CBR Based on our novel Breast Cancer ontology model, we propose ontologies4 -considered interesting perspective in CBIR to bridge the semantic gap –also for Case-Based Reasoning, to structure, represent and incrementally learn knowledge contained in the cases. Results We currently evaluate our approach at the indexing level by developing an automated semantic indexing method. Individually, the most accurate results were obtained for the mitosis count. Although, a 7, 33 % error was registered for the training dataset and 11% for the testing dataset in a local grading, for the global grading we achieved no computation errors. Compared with the manual grading given by the pathologists, we achieved an accuracy of 80% for the breast cancer global grading. Conclusions

The scope of this paper is to firstly set up a theoretical foundation for this CBIR and CBR generic and systematic overview to emphasize the need of TISBR approach. The core distinctions between CBIR and CBR are presented in Table5: Content-Based Image Retrieval (CBIR)  image based  limited to retrieval phase  query expansion  lack of knowledge injection  semantic web-based  structured by image only  week learning/static database  context-dependent  open loop : naïve relevance feedback

Case-Based Reasoning (CBR)  textual information based  integrated new case after adaptation & revise  no query  a priori knowledge  knowledge-based  structured by cases  incrementally learning/dynamic database  context-modeling  close loop: case adaptation- case storage

Table5. General conclusions on CBIR & CBR

With an accuracy of 80% for the indexing, TISBR launches interesting perspectives for semantic retrieval and query refinement as complex reasoning strategies in medical research, mapped to specific applications. Some References: 1 A. Smeulders, M.Worring, S.Santini, A.Gupta, R.Jain, “Content -Based at the End of the Early Years”, IEEE Transactions on Pattern Analysis and Machine Intelligence, vol.22, no.12, pp. 1349- 1380, 2000 2 F.Long, H.Zhang, D.Feng, “Chapter 1- Fundamentals of Content- Based Image Retrieval”, Multimedia Information Retrieval and Management – Technological Fundamentals and Applications, David Dagan Feng, W. C. Siu & Hongjing Zhang (eds), Springer-Verlag, Germany, pp.1-26, 2003 3 H. Müller, N. Michoux, D.Bandon, A. Geissbuhler “A Review of Content-Based Image Retrieval System in Medical Applications- Clinical Benefits and Future Directions”, International Journal of Medical Informatics, vol.73, pp. 1-23, 2004 4

B.Smith, “Beyond Concepts: Ontology as Reality Representation”, Proceedings of FOIS 2004, International Conference on Formal Ontology and Information Systems, Achille Varzi and Laure Vieu (eds.), Turin, pp. 112, 2004

5

N. Vasconcelos, “From Pixels to Semantic Spaces: Advances in Content-Based Image Retrieval”, Computer, vol.40, no.7, pp. 20 -26, 2007

6

T.Deserno, S. Antani, R.Long, “Gaps in content-based image retrieval”, Proceedings of SPIE-The International Society for Optical Engineering, vol.6516, 65160J, pp.1-11, 2007 R.Datta, D.Joshi, J.Li, J.Wang, “Image Retrieval: Ideas, Influences, and Trends of New Age”, ACM Transactions on Computing Surveys, vol.40, nr.2, pp.1-66, 2008 A.Tutac, D.Racoceanu, T.Putti, W.Xiong, W-K.Leow, V.Cretu, “Knowledge-Guided Semantic Indexing of Breast Cancer Histopathology Images”, BioMedical Engineering and Informatics: New Development and the Future, Proceedings of the First International Conference on BioMedical Engineering and Informatics, Yonghong Peng and Yufeng Zhang (eds), IEEE Computer Society Publisher, vol.2, pp. 107-112, Hainan, 2008 J.Kolodner, “Case-Based Reasoning”, Morgan Kaufmann Publishers, San Francisco, 1993 A.Aamodt, E.Plaza, “Case Based Reasoning: Foundational Issues, Methodological Variations and System Approaches”, AI Communications, vol.7, no.1, pp.39-59, 1994 I.Watson. F.Marir, “Case-Base Reasoning: A review”, The Knowledge Engineering Review, vol.9, nr.4, pp. 355-381, 1994 G. Kamp, S. Lange, C. Globig, “Case-based reasoning technology: related areas”, Case-based Reasoning Technology: from Foundations to Application, M. Lenz (ed.), LNAI no. 1400 Springer, Berlin, pp. 327, 1998 I.Watson, “Case based reasoning is a methodology not a technology”, Knowledge-Based Systems, vol.12, pp. 303- 308, 1999 M.Nillson, M.Sollenborn, “Advancements and Trends in Medical Case-Based Reasoning: An overview of Sistems and System Development”, pp1-6, 2004

7 8

9 10 11 12 13 14

Remark: the semantic indexing approach was presented to BMEI 2008 conference, as a solution for the semantic gap in a CBIR system. An updated paper – with rules improved, ontology updated and validated- was recently accepted by MIAAB workshop of MICCAI 2008. The novelty of this current work consists of a detailed analysis on CBIR versus CBR for launching a hybrid strategy.