1. In Search of the Optimal Set of Indicators when Classifying Histopathological Images
Catalin Stoean
University of Craiova, Romania

2. Histopathological images
Histopathology represents the microscopic examination of biopsies in order to locate and classify disease
Clinical diagnosis of cancer, identification of the malignancy level: study of histological images
Pathologists search for regularities of cell shapes and change in distribution of the cells across the tissue.
Based on their expertise and personal experience, they decide whether the examined tissue regions are cancerous and also determine the malignancy level.
Histology is the study of the microscopic anatomy of cells and tissues of organisms
Histological analysis represents the examining of a thin slice of tissue under microscope
80% of the 1 million prostate biopsies yearly performed in U.S. are benign so, a pathologist spends only 20% of their time grading the malignant tissues. The distribution is not equally useful in the colorectal cancer cases, but still a significant amount of work could be saved for the pathologists

3. Histopathological images
The judgment of the pathologists, although educated and sometimes based on vast experience, is subjective and often leads to considerable variability
Variation in the classification of the same image by the same pathologist at different times (intra-observer) and the variation in the classification of the same image by different pathologists (inter-observer)
Quantitative image-based evaluation is necessary for reducing or eliminating inter- and intra- observer variations in diagnosis

4. Histological images – healthy tissues, 10x

5. Histological images – grade 1, 10x

6. Histological images – grade 2, 10x

7. Histological images – grade 3, 10x

8. Normal tissue
Grade 1
Grade 2
Grade 3

9. The main computational steps on histopathological images
The tissue sample is first dissected and fixed to preserve its structure.
It is then dehydrated, cleared and embedded.
Then, it is sectioned into very thin slices (e.g. 5 m), sections are mounted on a glass slide for staining via hematoxylin and eosin.

10. Feature extraction
Both, glands and nuclei are considered for the information retrieval process.
For glands¥:
Thresholding, watershed, erosion, dilation
For nuclei:
Grayscale -> normalized box filtering -> Thresholding -> Distance transform algorithm -> Image normalization -> Binary Image
¥ ACM GECCO 2015, Madrid

11. Feature extraction
Both glands and nuclei are considered for the information retrieval process.
The identified structures (glands, nuclei) are counted.
Delaunay triangles and Voronoi diagrams are obtained from the central points that enclose the contours of the detected structures.
For each structure area, the perimeter and the radius of the enclosing circle are computed.
For each measure, the mean, median, minimum/maximum and standard deviation are calculated.

12. Average, median, standard deviation, min/max
76 features extracted
Feature
Measures
Statistics
Total
Morphological
Area, perimeter, radius for glands
Average, median, standard deviation, min/max 12
Area, perimeter, radius for nuclei
Number of glands and of nuclei - 2
Topological
Area, perimeter, radius for the Delaunay triangles for glands
Area, perimeter, radius for the Voronoi polygons for glands
Area, perimeter, radius for the Delaunay triangles for nuclei
Area, perimeter, radius for the Voronoi polygons for nuclei
Number of Delaunay triangles for glands and nuclei

13. Data set & Classification
On the obtained data, support vector machines, linear kernel (proved better than the radial one)
76 indicators, 357 samples, all 800x600 pixels¥.
Previously*, also considered PCA and reduced the number of attributes from 76 to 13.
On this data, radial SVM – better than linear (by 4%).
Training data: 2/3 of entire data, 30 repeated runs.
SVM: 79.89%
SVM + PCA: 74.85%
The PCA reduced the number of attributes from 76 to only 13, as this was the number of components required to explain at least 95% of the variance.
* C. Stoean et al., SVM-Based Cancer Grading from Histopathological Images using
Morphological and Topological Features of Glands and Nuclei, Intelligent Interactive
Multimedia Systems and Services, 55, Smart Innovation, Systems and Technologies,
Springer, pp , 2016.
¥Dataset available for download at

14. Feature selection
Two filter methods are tried (via the RSelector R package):
Consistency-based filter
Correlation mechanism (CFS)
One wrapper method – GA (via the genalg R package)
binary representation
Individual size: 76
Fitness function: SVM with linear/radial kernel, 2/3 training and 1/3 test data, 30 repeats.
GA parameter tuning:
the choices comprised a population size of 50 and 100, a number of iterations to give the stop condition between 30, 40 and 50 and the mutation probability in {0.005,
0.01, 0.1}. During the parameter tuning phase, the number of repeated runs of the SVM was reduced to only 10 and the accuracy was computed as an average of the 10 classification
accuracies.
Fitness
Population size
Iterations
Mutation probability
SVM linear
100 50
0.01
SVM radial
0.1

15. Results CFS selects only 3 features. Consistency appoints 15 features.
GA & SVM linear: 20 features.
GA & SVM radial: 24 features.
Classifier GA
CFS
Consistency
SVM linear
83.94%
49.59%
69.97%
SVM radial
81.9%
53.38%
77.08%

16. Most selected features by the GA
SVM Linear kernel
SVM Radial kernel

17. Most selected features
34 features are not selected by any of the methods

18. Conclusions & Future Work
A data set of 357 histopathological images, 800x600 pixels, separated into 4 grades.
Diagnosis accuracy over the four classes: 83.94%
More options in the computational steps remain to be tried
Parameter tuning
Only thresholding was previously considered

19. Q&A