Resumen de Computer aided breast cancer detection and diagnosis system based on histopathological image analysis of tmas
María del Milagro Fernández Carrobles
Currently, TMA analysis is performed manually by the pathologist who gives a diagnostic based on microscopic observations of the biopsies samples. In this way, the evaluation can be subjective. For that reason, the automation of this task is fundamental in order to provide the pathologist with a tool which automatically analyses (at different magnifications) the samples and produces a diagnostic result. TMAs can gather dozens or even hundreds of tumours in a paraffin block and be used to analyze large molecular markers. Besides, it should be pointed out that TMAs allow carrying out simultaneous and standardized studies of multiple samples with uniform staining. All this allows reducing the economical and temporal cost of TMA preparation and interpretation.
breast TMA which was able to automatically acquire and classify TMA cores. For that purpose, in the acquisition process several image processing algorithms were created and applied on the TMA thumbnail image to detect, select and archive the cores in an individual image at different magnifications, that is, 5x, 10x, 20x and 40x. On the other hand, the tissue examination and classification process were considered. In this regard, the author conducted a thorough analysis of the existing techniques in histopathological image classification. This preliminary analysis led to a complete study of breast tissue based on colour and texture features. TMA cores extracted from the previous core acquisition algorithm were used to select the 628 ROIs of 4 representative tissue classes. The size of these regions was 200 x 200 pixels (0.74_m/pixel at 10x) and the TMA tissue classes were categorized as: i) benign stromal tissue with low and medium cellularity (170 images), ii) adipose tissue (103 images), iii) benign structures and anomalous (163 images) and iv) different kinds of malignity, that is, ductal and lobular carcinomas (192 images). Once the dataset was established it was first transformed and then filtered by several colour models and texture descriptors respectively.
Therefore, a relevant set of features was obtained on 8 different colour models: RGB, CMYK, HSV, Lab, Luv, SCT, Lb and Hb from from 1st and 2nd order statistical descriptors obtained from the intensity image, Fourier, wavelets, Gabor, M-LBP and texton descriptors. The number of features extracted depends on the descriptor used. Fourier, wavelets and Gabor descriptors have four bands to build their filtered images so that their filtered images are multiplied by 4. Finally, intensity, M-LBP and texton (frequential and spatial) statistic descriptors contain and average of 229 features for each colour model and Fourier, wavelets and Gabor statistic descriptors contain and average of 916 (229x4) features. However, the large number of features may produce redundant information and increase the computational time. Therefore, a dimensionality reduction of the feature sets is needed. In this PhD thesis three methods of dimensionality reduction were analysed: linear discriminant analysis, correlation and sequential forward search.
Finally, the training and classification stages are applied, using 10-fold cross-validation and 5 different classifiers: Fisher, support vector machine (SVM), Bagging, random forest and AdaBoost.
The results obtained in both core acquisition and classification were highly satisfactory.
For core acquisition, 4 TMA datasets with a total of 21244 cores had been processed obtaining an average of 98% accuracy. For TMA core classification four types of classification experiments were performed, that is, (1) Classification per colour model individually, (2) Classification by combination of colour models (3) Classification by combination of colour models and descriptors and (4) Classification by combination of colour models and descriptors with a previous feature set reduction. The best result in classification experiments was obtained with the CMYK&Hb&Lb&HSV&Luv&SCT colour combination and Intensity&M-LBP&Gabor&Spatial Textons descriptors reaching an average of 99.045% accuracy and 98.34% precision with a total of 1719 features. Once all the algorithms were developed they were integrated into a complete CAD tool called TMA CAD System.
In conclusion, CAD systems in histopathology are still a challenge due to the fact that histopathological images encompass a variety of cancer types and the analysis is still performed by the pathologist under the microscope. It is therefore important the development of applications such as the TMA CAD System developed in this PhD thesis.
Firstly, the tool allows to acquire the TMA cores individually which is essential for further tissue classification and secondly there is an improvement in breast tissue classification by combining colour and texture descriptors. Pathologists at the department of Anatomical Pathology of Ciudad Real have tested and evaluated the TMA CAD System. They assessed the tool very positively.
