In this article we present SDCA, which is a system to detect cancerous abnormalities in digital mammograms. The SDCA try to give at radiologist a second opinion in the analysis of a digital mammogram to increase the reliability of detecting breast cancer. SDCA is a semiautomation of KDD process (Knowledge Discovery in Databases). The KDD process is a method that uses strategies of Artificial Intelligence (AI) to extract patterns of behavior in databases with large volumes of information. Two SDCA characteristics outstanding are 1) the implementation of Mej´ıa filtering method in the data cleansing module, and 2) the implementation of Decorate strategy Classification in the classification module. The results shows that SDCA get 95% of detections classified correctly. SDCA was developed using Matlab GUIDE, and tests were done with the database (DB) of digital mammographic MIAS.
For the radiologist, mammograms are highly useful to identify abnormalities carcinogenic potential. The difficulty arises when the radiologist’s review does not guarantee the detection of cancerous abnormalities. Therefore, this research serves as a support in the detection of abnormal regions. Therefore, the area of interest for this research is the analysis of medical images using the KDD process.
This research, then, serves to give the radiologist a second opinion on the detection of abnormal and thus increase the reliability of diagnosis.
SDCA is an extension of work previously presented in [
The rest of the paper is divided as follows: Section 2 briefly describes the KDD process for this research. Section 3 describes the filter module. Section 4 describes the segmentation module. Section 5 describes the classification module. Section shows experimental results and finally in section ends with a conclusion and future work. 2
The following describes the KDD process steps for this research:
– Selection: Given a set of different digital mammography DB, the most
representative was selected respect its use in other research. MIAS (see Table
1) is a reduced version of the original, and for a long time, been used to
test research in [
– Data Preparation: The images are filtered using the Mej´ıa filtering method, to reduce noise. – Data Transformation: Segmented manually, the area of interest by the radiologist, normalizes each segmented area to have the same dimensions and finally gives the frequency histogram of gray levels of each segmented area, and stored for create a testing base. – Data Mining: new samples are obtained and applied Decorate classification strategy to classify the abnormality, if it exists.
– Patterns evaluation: The patterns obtained are evaluated by the radiologist. 3
This module was integrated into the prototype, because with this filtering method
is obtained excellent results in the enhancement of abnormalities. This work was
presented in [
In this module, the radiologist manually segmented the area he considers abnormalities. The result is shown in Fig. 2. Then the area of interest is normalized so that all images have the same dimensions.
After being normalized interest area, gives the frequency histogram of gray levels, ranging from 0 to 255 (see Fig. 3). This histogram is stored in a xlsx file called Histograma.xlsx for purpose of building the testing base.
Then the radiologist choose a tag if the detected abnormality is benign (B), malignant (M) or normal (N). Finally for this module, open the file Histograma.xlsx and saved in CSV format (comma delimited) for later use. 5
In this module selects the Histograma.csv file to implement the classification of the abnormality: The classification is activated by pressing the Detecci´on button, which indicates the call to Decorate strategy, which uses, in this case, the algorithm LADTree base, and this, in turn, is based on the algorithm LoogitBoost 1.1.
Result: If p (1|a)>0.5 predict the first class Else the second 1 begin 2 for j = 1 until t do 3 for a [i] do 4 Assign the target value for the regression to z [i] = (y [i] − p (1 |a [i])) (p (1 |a [i])Λ (1 − p (1 |a [i])) Assign the weight of the instance to [i] = p (1 |a [i])Λ (1 − p (1 |a [i])) 5 Fit a regression model fj at data with class values z [i] and weights w [i] 6 7 8 end
Finally, the result is released (see Fig. 6) for analysis by the radiologist. This table shows four predictions made by SDCA, the first three samples are needed for the proper operation of the system, however, the new sample, which interests the radiologist is the fourth, and in this new sample the radiologist predicts the area of interest was the first type, i.e., benign (1: B) (blue circle), but the system indicates that normality can be either three, or malignant (3: M) (red circle). This aid increase the reliability, as is now required to take another mammogram from another angle to confirm whether the abnormality is type 3: M. 6
Classification Algorithms We tested each classification algorithm for each strategy, the strategies are: bayesian algorithms (bayes) classification functions (functions), algorithms to generate rules (rules), meta classifiers (meta), lazy algorithms (lazy ), algorithms generation of decision trees (trees) and miscellaneous algorithms (misc).
We used the 322 digitized mammograms of MIAS BD (Table 1) for the first test and even to choose the algorithm that performed better. As shown in the table below, the best result was obtained with the strategy Decorate. This strategy was 95% of instances correctly classified.
Strategy Algorithm CCI* % of CCI* Bayes NaiveBayes 13 65% Functions Logistic 14 70% Rules PART 12 60% Meta Decorate 19 95% Lazy IB1 12 60% Trees RandomForest 17 85%
Misc HyperPipes 10 50%
* CCI ← Correctly Classified Instances
Experiments and analysis For testing we used a data sample of eighty,
divided into four sets of twenty instances each. These results have been presented
previously [
In the first dataset is obtained 95%; in the second dataset is obtained, similarly, 95%; in the third dataset is obtained 90% and finally, in the fourth dataset is obtained 100% of instances correctly classified. These results fluctuate between 90% and 100%, giving an average of 95% of instances correctly classified. This shows that SDAC is very reliable to use in the classification of cancerous abnormalities. Evaluaci´on. En los u´ltimos an˜os se han presentado aportaciones para ayudar al radio´logo en el diagno´stico de c´ancer de mama. Las pruebas en estos trabajos se han realizado con diferentes BDs.
En la tabla 2, vemos que [
SDAC
[
En este trabajo, presentamos SDCA para detectar anormalilades cancer´ıgenas en mastograf´ıas digitales. SDCA es una semi-automaizaci´on del proceso KDD. Los resultados obtenidos muestran que SDCA aumenta la fiabilidad en la detecci´on de anormalidades cancer´ıgenas, dando al radio´logo una segunda opinio´n sobre la revisio´n de la mastograf´ıa.
Como trabajo futuro se propone aumentar el nu´mero de datos de pruebas, para
verificar que se mantenga el promedio de fiabilidad alrededor de 95%. Adema´s,
teniendo en cuenta los buenos resultados aqui obtenidos e inspirados en [