Optical biopsy

IntroductionMaterials and Methods

ResultsDiscussion

Automatic Classification of Optical BiopsyImages

Ruben Tous1 & Prof. Dr. Ferrer Roca

2

1Universitat Politècnica de Catalunya. Departament d’Arquitectura de

Computadors

2 Unesco Chair of Telemedicine. Canary Islands. Spain

July 2013

Tous Automatic Classification of Optical Biopsy Images


ResultsDiscussion

Índex

1 Introduction

2 Materials and Methods

3 Results

4 Discussion



ResultsDiscussion

Índex

1 Introduction


3 Results

4 Discussion



ResultsDiscussion

Introduction

Problem statementConfocal laser endomicroscopy (CLE) has revolutionizedgastrointestinal endoscopy by providing microscopicvisualization on a cellular basis in vivo.However, most gastroenterologists are not trained tointerpret mucosal pathology, and histopathologists areusually not available in the endoscopy suite.



ResultsDiscussion

Introduction

General project goal

To build a set of computer-aided diagnosis tools that assistendoscopists in the interpretation of optical biopsies obtainedthrough CLE.

Real-time functionalities for supporting diagnosis in vivo.Tools for cataloguing and searching CLE databases.



ResultsDiscussion

Introduction

Project subgoals

Real-time functionalities for supporting diagnosis in vivo.Automated diagnosis. (normal vs. abnormal mucosa,differentiation between different pathologies).Overprint visual markers highlighting the geometry of thecrypts.

Tools for cataloguing and searching CLE databases.Automated cataloguing of CLE databases.Search engine with query-by-example functionality.



ResultsDiscussion

Introduction. Colon microarchitecture



ResultsDiscussion

Introduction. Colouring agents

Left: acrifliavine hydrochloride. Right: fluorescein (used for oursample). Red arrows: goblet cells.



ResultsDiscussion

Índex

1 Introduction


3 Results

4 Discussion



ResultsDiscussion

Materials and Methods. General workflow



ResultsDiscussion

Materials and Methods. General workflow

In a first step, the system’s database was standardizedusing image sets based on the Mainz confocalclassification for colonic optical biopsies.This step was used to provide an optical biopsy searchengine based on extracting simplified features of crypt andmucosal patterns.The second step aims at providing an interrogationplatform to provide on-site assistance during CLE, mainlyby automated detection of crypt geometry.



ResultsDiscussion

Materials and Methods. Image set description

137 black and white instances10 are labelled as "healthy"12 as "hyperplastic"115 as neoplastic"

Instances have 1024x1024 pixels, and are 8-bit grayscale,JPEG compressed and 548KB in average.Many of the instances were taken from same patients, andsome of them only capture slight variations of a sametissue.



ResultsDiscussion

Materials and Methods. Class description

Each image instance is tagged according to medicaljudgement, the three main categories being:

HealthyHyperplasticNeoplastic

Neoplastic comprises Adenomous and Cancerous, each ofthem in several degrees of seriousness and malignity,namely normalör low gradefor adenomas, and normaländType 1for carcinomas.Sample instances can be categorised via visual inspection,at least into the three main categories.



ResultsDiscussion

Materials and Methods. Healthy tissue

Healthy tissue is characterised by regularcrypts in a honeycomb-like structure thatpresent small size and shape variance.In general, healthy crypts are recognisableas darker, circle-shaped areas in agenerally regular spatial distribution.Due to variations in the depth or diferentdegrees of fluorescent agent penetration,the cells’ cytoplasm may be especiallyprominent (bottom). This circumstancehas proven, at later stages, to benon-trivial to address.



ResultsDiscussion

Materials and Methods. Hyperplastic tissue

Hyperplasia means cells have proliferatedmore than usual in a tissue.In hyperplastic colon cases we canappreciate abnormal crypt bloating,although their integrity is still preserved tosome extent.



ResultsDiscussion

Materials and Methods. Hyperplastic tissue

Some milder cases present a regular cryptstructure just like in healthy tissue, exceptin one or two crypts that have abnormalgrowth.In contrast with these, some others aremore dificult to recognise.



ResultsDiscussion

Materials and Methods. Neoplastic tissue

Neoplasia is the massive and uncoordinated proliferation ofcells; this is also commonly known as tumour.



ResultsDiscussion

Materials and Methods. Neoplastic tissue

A typical feature in them is having strangenuclear-to-cytoplasmic ratios. However,nuclei are not stained by fluorescein.Fortunately, neoangiogenesis can beobserved with fluorescein, and it is aphenomenon whose detection pointsstrongly towards pathogenesis.These new blood vessels are caracterisednot only by tortuosity and high irregularity,but also by profuse liquid leakiness, thatcauses tissue staining easy to see.



ResultsDiscussion

Materials and Methods. Algorithm outline

Automatic feature extraction in CLE images. Inferringsemantic metadata from low-level features.Image Analysis + Pattern Recognition.The image analysis algorithm allows identifying thedifferent crypts and also featuring their contours.The extracted low-level visual features are then combinedto obtain a feature vector.This vector is analyzed to translate the low-level detailsinto high-level semantic information about the images,notably a suggested diagnosis.



ResultsDiscussion

Materials and Methods. Algorithm flowchart



ResultsDiscussion

Índex

1 Introduction


3 Results

4 Discussion



ResultsDiscussion

Results. Automatic crypt geometry characterization

The image analysis algorithm allows identifying thedifferent crypts and also featuring their contours with highaccuracy.This information is computed to characterize the geometryof the crypts and to overprint visual markers aiming tofacilitate diagnosis.



ResultsDiscussion

Results. Automatic crypt geometry characterization

Two watershedding flavours on aneoplastic image



ResultsDiscussion

Results. Automated diagnosis

Our classification approachChaining two binary classification problems in a row. 1)healthy vs. unhealthy; if unhealthy 2) hyperplasia vsneoplasia.We opted for this because the crypt shape, boundary, etcon hyperplastic examples were not much different fromthose obtained from neoplastic ones.All classifers were evaluated using Leave One OutCross-Validation (LOOCV).



ResultsDiscussion


(step 1) healthy vs. unhealthyEffectiveness (with 200x200 pixels downscaling and acustom classifier). A 100% hit rate is attained with thepresent algorithm.The present sample (137 images) is relatively small, andits classes are highly unbalanced in size.Besides, they come from an even smaller number ofpatients and capturing situations.



ResultsDiscussion


Confusion Matrix (custom classifer healthy vs. unhealthy)

classified as unhealthy classified as healthy

unhealthy 127 0healthy 0 10



ResultsDiscussion


(step 2) hyperplasia vs neoplasiaClassfication over the resulting unhealthy examples fromstep 1.We noted that cells were usually abundant andhomogeneously distributed in hyperplastic images.We implemented a simple adhoc decision tree on the only3 descriptors for the second decision by seeing how theybehaved. It also reached 100% prediction rate withinunhealthy images when dividing them in hyperplasia andneoplasia.



ResultsDiscussion


EfficiencyWe have studied the impact of the parameters choice ontime costs, aiming to reduce them at the least possibleaccuracy expense.With the current performance, we believe real-timeanalysis of the newly acquired images is possible, thusaiding the endoscopist in his or her exploration task.In addition, this suggests a possible future line of progress:work towards a video-rate (24 frames per second)processing system, where crypt highlighting could beespecially valuable.



ResultsDiscussion

Results. Cataloguing and search

Besides supporting the diagnosis of a single image duringan ongoing session, the described method also allowsannotating a complete database of CLE images withsemantic information, thus providing a search engine withadvanced functionalities such as semantic retrieval orquery by image.



ResultsDiscussion

Índex

1 Introduction


3 Results

4 Discussion



ResultsDiscussion

Discussion

On-site image analysis and semantic interpretation areable to provide a standardized diagnosis of CLE images inreal time.This has the potential to facilitate, standardize and shortenendoscopists’ training for CLE.In addition, we provide a powerful tool to exploredatabases of images for retrospective analyses.This system’s geometry is not limited to CLE, but can beextended to multiple imaging modalities.


Optical biopsy

Technology

introduction materials

optical biopsy search

colonic optical biopsies

mainz confocal classication

automated diagnosis

aided diagnosis tools

gastrointestinal endoscopy

endoscopy suite