TOWARD A DIAGNOSIS ASSISTANCE SYSTEM FOR DIGITAL PATHOLOGY

TOWARD A DIAGNOSIS ASSISTANCE SYSTEM FOR DIGITAL PATHOLOGY OF BREAST CANCER M. Oger, P. Belhomme, J. J. Michels, A. Elmoataz GRECAN, EA 1772, University of Caen Basse-Normandie F. BACLESSE Cancer Centre, Caen GREYC, UMR 6072, University of Caen Basse-Normandie Digital Pathology Solutions Conference m. oger@baclesse. fr

Introduction • Identification of breast tumor lesions is not always a easy task. • Cancer lesions are sometimes heterogeneous. • Question: is automatic image processing able to help classifying benign and malignant breast lesions? Digital Pathology Solutions Conference m. oger@baclesse. fr

Example Digital Pathology Solutions Conference m. oger@baclesse. fr 237 Mb

Aim • To try to develop automated Computer-Aided Diagnosis (CAD) tools for pathologists • To work with Virtual Slides (VS) in order to take into account lesion heterogeneity Digital Pathology Solutions Conference m. oger@baclesse. fr

Material and method • Low resolution Virtual Slide 6 µm: Nikon Cool. Scan 8000 ED. slide holder • 224 images (different size) are included in the knowledge base • 28 histological types • 3 histological families (Benign, Malignant Carcinoma, Malignant Sarcoma) images with foci of different histological type exist, but we labeled them according to the dominant type Digital Pathology Solutions Conference m. oger@baclesse. fr

Example of low resolution VS Fibroadenoma 3479 X 2781 px = 28 Mb Intraductal carcinoma 2228 X 1915 px = 12. 3 Mb • At the resolution of 6 µm, pathologists recognize fairly easily histological types in 80 to 90% of cases. but “small objects” are sometimes difficult to identify Digital Pathology Solutions Conference m. oger@baclesse. fr

Material and method • A “new image” will be compared to the knowledge database. • A graphical user interface will be built to allow a “visual” presentation of the results obtained. Digital Pathology Solutions Conference m. oger@baclesse. fr

Strategy Exploration • Multiparametric Analysis CAD system 1 st version • Spectral Analysis CAD system 2 nd version Digital Pathology Solutions Conference m. oger@baclesse. fr

Multiparametric analysis • We have developed a system which statistically determines the “similarity degree” of a new image compared to the different histological types. • Requirements: » No segmentation » Exploration of several color spaces: RGB, YCH 1 CH 2 (Carron), AC 1 C 2 (Faugeras), I 1 I 2 I 3 (Ohta). . . • Application: » Computing a “signature” of parameters of the whole VS » Comparing the signatures Digital Pathology Solutions Conference m. oger@baclesse. fr

The color signatures • 234 global parameters computed on 6 color spaces – – – Histograms Mean Median Kurtosis Skewness… Principal Component Analysis 188 • + 13 "texture" parameters – S/N measure – Haralick… • Vector distance (comparison of signatures) – Kullback-Leibler distance • Software development – PYTHON language Digital Pathology Solutions Conference m. oger@baclesse. fr

CAD 1 st version system Automated system ► Input = a new image ► Outputs = similar images from the knowledge base ► Digital Pathology Solutions Conference m. oger@baclesse. fr

CAD 1 st version: Results Exhaustive analysis of the image database (one image vs the 223 others) with Kullback-Leibler distance Rank of the first image of the same type 1 13. 99 % ≤ 3 33. 33 % ≤ 5 47. 74 % ≤ 10 67. 08 % Digital Pathology Solutions Conference m. oger@baclesse. fr

Comments • Low resolution image classification is possible but this strategy is a crude one which can lead only to a “preclassification” of the lesion under study • Other strategies are to be explored Digital Pathology Solutions Conference m. oger@baclesse. fr

Strategy Exploration • Multiparametric Analysis CAD system 1 st version • Spectral Analysis CAD system 2 nd version Digital Pathology Solutions Conference m. oger@baclesse. fr

Principle of spectral techniques for structural analysis of an image database • Working on images with identical size • Comparing “point to point” each image with all those of the database ==> the signature is the WHOLE image • Trying to determine a “distance” between all the images of the database by using techniques of Spectral Dimensionality Reduction • Replacing a n-dimensional space by a 2 D-visualization space (φ1, φ2) Digital Pathology Solutions Conference m. oger@baclesse. fr

Application to breast lesions • Problem: – Database images are of various size – In an image, some areas are uninformative (stroma, normal tissue, adipose cells. . . ) • Proposed solution: – Finding the interesting “PATCHES” which describe the histological type at best – Choosing an adequate size for “patches”: 32 x 32 px² Digital Pathology Solutions Conference m. oger@baclesse. fr

Example of 4 distinct classes Intra Ductal Carcinoma Invasive Lobular Carcinoma • We work with: – – Intra Ductal Carcinoma Invasive Lobular Carcinoma Colloid Carcinoma Fibroadenoma • We take only the 3 most representative VS of each class(□) 12 VS among 73 Colloid Carcinoma Fibroadenoma Digital Pathology Solutions Conference m. oger@baclesse. fr

250 patches from each VS FA IDC CC ILC Digital Pathology Solutions Conference m. oger@baclesse. fr 250 x 3 x 4 = 3000 retained patches

Graph of the selected 4 types Colloid Carcinoma Fibroadenoma Intra Ductal Carcinoma Invasive Lobular Carcinoma Digital Pathology Solutions Conference 1 cross per patch = 3000 crosses m. oger@baclesse. fr

How can we analyse a “new image” • 1) elimination of the background Digital Pathology Solutions Conference m. oger@baclesse. fr

• 2) Cutting in 32 x 32 patches Digital Pathology Solutions Conference m. oger@baclesse. fr

• 3) « patches » are projected on a 2 D space (φ1, φ2) Digital Pathology Solutions Conference φ1 = 0 m. oger@baclesse. fr

• 4) segmentation by spectral analysis: patches corresponding to stroma are removed (cellular zones are preserved) Stroma Cellular zones Digital Pathology Solutions Conference φ1 = 0 m. oger@baclesse. fr

Visual control • 4) segmentation by spectral analysis: patches corresponding to stroma (Green) are removed, cellular zones (Purple) are preserved Digital Pathology Solutions Conference m. oger@baclesse. fr

CAD 2 nd version • 5) cellular patches of the new image are projected onto the graph of cellular patches of the 4 histological types Digital Pathology Solutions Conference m. oger@baclesse. fr Insertion of the new image

CAD 2 nd version Results of a test done with a “new image” corresponding to an Intraductal Carcinoma Matching probabilities 2 -neighborhood k-neighborhood Intra Ductal Carcinoma 42, 37% Invasive Lobular Carcinoma 5, 64% Colloid Carcinoma 29, 98% Fibroadenoma 22, 01% Digital Pathology Solutions Conference m. oger@baclesse. fr Detail of the whole graph

Conclusion • Technique of spectral analysis seems to be promising regarding 4 classes of tumors. • This technique could be applied in order to try to identify tumor foci of different types on a virtual slide. Digital Pathology Solutions Conference m. oger@baclesse. fr

Perspectives • But a lot of work remains to be done: – Extending the spectral analysis to 28 classes (the rest of the database): improving the separation of the influence zone of each histological type. – Increasing the signature: image patch + parameters which have been selected in the first part. – Testing a higher resolution (sub sampled high resolution virtual slides). Remark: the final strategy will be easily applicable to other tumor locations Digital Pathology Solutions Conference m. oger@baclesse. fr

Acknowledgements: The authors gratefully acknowledge Dr Paulette Herlin, Dr Benoît Plancoulaine, Dr Jacques Chasle, the Regional Council of "Basse-Normandie" and the "Comité départemental du Calvados de la Ligue de Lutte Contre le Cancer". Digital Pathology Solutions Conference m. oger@baclesse. fr