UMR • CNRS • 5516 • SAINT-ETIENNE Automatic Nonlinear Filtering and Segmentation for Breast Ultrasound Images Mohamed Elawady, Ibrahim Sadek, Abd El Rahman Shabayek, Gerard Pons, Sergi Ganau P ROBLEM D EFINITION The ultrasound image contrast between the ab- normality and the surrounding breast tissue is insufficient for direct lesion detection. C ONTRIBUTION • Introducing a fully automatic lesion extraction algorithm. • Proposing a fast segmentation step by means of Quick Shift; against frequently used Normalized Cut. • Conducting a comparative study on the most common preprocessing nonlinear techniques. M ETHOD Figure 1: The proposed framework used for lesion segmentation in BUS images. R ESULTS I Performance results across all proposed methods (segmentation [QS: Quick Shift and NC: Normal- ized Cut] and preprocessing [FR: Frost Filter, DPAD: Detail Preserving Anisotropic Diffusion and PPB: Probabilistic Patch-Based]): Percentage (%) 0 10 20 30 40 50 60 70 Methods QS-FR QS-DPAD QS-PPB NC-FR NC-DPAD NC-PPB Dice Jaccard Sensitivity Figure 2: Statistical metrics calculated in average. Figure 3: Box plot of Dice similarity coefficient. RESULTS II Results of some successful lesion extraction. First row represents some of the input images. Second, third and fourth rows show the output results of [QS-FR, NC-DPAD, NC-PPB] respec- tively, in which white color is true segmented lesion, green color is false positive, red color is false negative and black color is true negative. The computation time of Quick Shift method is 8x faster than Normalized Cut method. The failure cases exist in all methods due to the intensity similarity of surrounding tissues around the target lesion, leading to incorrect segmentation. R EFERENCES [1] H.D. Cheng, Juan Shan, Wen Ju, Yanhui Guo, and Ling Zhang. Automated breast cancer detection and classification using ultrasound images: A survey. Pattern Recognition, 43(1):299 – 317, 2010. [2] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(8):888–905, August 2000. [3] A. Vedaldi and S. Soatto. Quick shift and kernel methods for mode seeking. In European Conference on Computer Vision, 2008. [4] Ju Zhang, Chen Wang, and Yun Cheng. Comparison of despeckle filters for breast ultrasound images. Circuits, Systems, and Signal Processing, 34(1):185–208, 2015. C ONCLUSION • Best performance: FR with QS, DPAD with NC and PPB with NC. • QS is a more preferable choice in real time applications. • Future work: use superpixel segmentation approaches for robust results. P ARTNERS

ICIAR 2016 Poster: Automatic Nonlinear Filtering and Segmentation for Breast Ultrasound Images

Download PDF Report

Upload
universite-jean-monnet-saint-etienne
View
221
Download
1

Embed Size (px)

Citation preview

UMR • CNRS • 5516 • SAINT-ETIENNE

Automatic Nonlinear Filtering and Segmentationfor Breast Ultrasound Images

Mohamed Elawady, Ibrahim Sadek, Abd El RahmanShabayek, Gerard Pons, Sergi Ganau

PROBLEM DEFINITION

The ultrasound image contrast between the ab-normality and the surrounding breast tissue isinsufficient for direct lesion detection.

CONTRIBUTION

• Introducing a fully automatic lesion ex-traction algorithm.

• Proposing a fast segmentation step bymeans of Quick Shift; against frequentlyused Normalized Cut.

• Conducting a comparative study on themost common preprocessing nonlineartechniques.

METHOD

Figure 1: The proposed framework used for lesion segmentation in BUS images.

RESULTS IPerformance results across all proposed methods (segmentation [QS: Quick Shift and NC: Normal-ized Cut] and preprocessing [FR: Frost Filter, DPAD: Detail Preserving Anisotropic Diffusion andPPB: Probabilistic Patch-Based]):

Perc

enta

ge (%

)

MethodsQS-FR QS-DPAD QS-PPB NC-FR NC-DPAD NC-PPB

DiceJaccardSensit ivity

Figure 2: Statistical metrics calculated in average. Figure 3: Box plot of Dice similarity coefficient.

RESULTS IIResults of some successful lesion extraction.First row represents some of the input images.Second, third and fourth rows show the outputresults of [QS-FR, NC-DPAD, NC-PPB] respec-tively, in which white color is true segmentedlesion, green color is false positive, red color isfalse negative and black color is true negative.

The computation time of Quick Shift methodis 8x faster than Normalized Cut method.The failure cases exist in all methods due tothe intensity similarity of surrounding tissuesaround the target lesion, leading to incorrectsegmentation.

REFERENCES

[1] H.D. Cheng, Juan Shan, Wen Ju, Yanhui Guo, and Ling Zhang. Automated breast cancer detection andclassification using ultrasound images: A survey. Pattern Recognition, 43(1):299 – 317, 2010.

[2] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on PatternAnalysis and Machine Intelligence, 22(8):888–905, August 2000.

[3] A. Vedaldi and S. Soatto. Quick shift and kernel methods for mode seeking. In European Conference onComputer Vision, 2008.

[4] Ju Zhang, Chen Wang, and Yun Cheng. Comparison of despeckle filters for breast ultrasound images.Circuits, Systems, and Signal Processing, 34(1):185–208, 2015.

CONCLUSION

• Best performance: FR with QS, DPADwith NC and PPB with NC.

• QS is a more preferable choice in real timeapplications.

• Future work: use superpixel segmenta-tion approaches for robust results.

PARTNERS