A Region of Interest Approach For Medical Image Compression Salih

A Region of Interest Approach For Medical Image Compression Salih Burak Gokturk Stanford University

OVERVIEW Motivation Previous Work Comparison Study of Compression Schemes ROI based System Design Conclusion

Motivation Medical images are huge.(300x512x512x2) High quality imaging is required in diagnostically important regions. ROI based approach is the only solution: – Lossless compression in ROI. – Very lossy compression in non-ROI.

OVERVIEW Motivation Previous Work Comparison Study of Compression Schemes ROI based System Design Conclusion

Previous Work Lossless Compression Schemes (Takaya95, Assche00) DCT based Compression Schemes (Vlaciu95) PCA based Compression(Tao96) Wavelet Transformation(2D and 3D) (Baskurt93) ROI based coding (Cosman 94,95)

OVERVIEW Motivation Previous Work Comparison Study of Compression Schemes ROI based System Design Conclusion

Lossless Compression Entropy of images – 7.93bpp Predictive Coding – 5.9bpp Entropy of difference images – 5.76bpp

DCT Compression (1)

DCT Compression (2)

DCT Compression (3) Quantization Step Size 1 MSE in dB -11.7 -5.7 2 4 8 16 32 64 128 256 512 1024 0.34 6.26 11.9 17.1 21.8 25.7 29.3 32.6 35.9 Rate (without RLC) 5.74 (bpp) 4.97 4.09 3.20 2.34 1.57 0.96 0.55 0.31 0.16 0.09 Rate (with RLC) (bpp) 7.09 5.87 4.51 3.15 1.95 1.07 0.55 0.28 0.14 0.07 8.04

PCA Compression - Treat each image block as a vector Rate 0.54 bpp MSE 30 dB

Blockwise Vector Quantization(1) - A simpler decoder is required

Blockwise Vector Quantization(2) MSE 38 dB MSE 39 dB

Motion Compensated Hybrid Coding (1) - Lukas Kanade Tracker was used by 0.1 pixel accuracy

Lukas-Kanade Tracker

Motion Compensated Hybrid Coding (2) - Entropy of the motion vector is 2.28 and 2.45 in x and y. - This brings 0.018 bpp. MSE 35 dB

OVERVIEW Motivation Previous Work Comparison Study of Compression Schemes ROI based System Design Conclusion

Segmentation - Thresholding to find the air - Gradient magnitude to extract the colon wall - Grassfire operation to find the ROI around the colon wall

ROI Based System

Experiment with 16 by 16 Blocks - The ratio of ROI %12.2 - Entropy of motion vector is 2.28 in x and 2.45 in y - The entropy of the error image is 4.38 - average RMS error 33.7 dB with lossless in ROI - Overall rate 0.552 bps MSE 33.7 dB

Experiment with 8 by 8 Blocks - The ratio of ROI %7.3 - Entropy of motion vector is 1.82 in x and 1.96 in y - The entropy of the error image is 4.31 - average RMS error 30.3 dB with lossless in ROI - Overall rate 0.37 bps MSE 33.7 dB MSE 30.3 dB

OVERVIEW Motivation Previous Work Comparison Study of Compression Schemes ROI based System Design Conclusion

Conclusion Effective System (compression rate of %2.3) Accurate System (lossless in ROI) Results of ROI based compression over performs standard compression schemes. Future work includes lossy compression in ROI. Case study with the radiologist for determining rate-diagnosis performance curve.