P6C-2 Image Texture Clustering for Prostate Ultrasound Diagnosis

To double the accuracy of ultrasound diagnosis and biopsy guidance, an efficient, integrated platform for image textural analysis and clustering of transrectal prostate ultrasound images into clusters potentially representing cancerous or normal tissue areas is presented. Preliminary image texture analysis has shown the potential for doubling diagnosis accuracy from 38-42% for prostate cancer with current clinical methods, to 88-92%. In addition, image texture analysis makes prostate cancer tumor locating possible for more precise, less invasive biopsy/treatment, instead of 6-way random biopsy. Due to the tedious derivation of co-occurrence matrix and scanning a sub-image tile (square image window) sequentially, pixel by pixel across and down the entire image, the initial image texture analysis on a NASA JSC´s miniVAX could take 8 days CPU time per image, i.e., more than 5 months analysis for 20 cross-sections per patient, making clinical point-ofcare diagnosis impossible. An efficient Image Texture Analysis tool platform on Window PC is constructed via innovative sparse co-occurrence matrix techniques with linked lists to speedup the processing from 8 days to about 5 seconds per image on a PC. The approach is based on Haralick ´s textural features [1] and the Mean Squared Error (MSE) clustering algorithm. This significant reduction in run time potentially allows more accurate, objective diagnoses to be performed within clinical settings, as well as enables the general investigation of image textural and clustering parameters. Ultrasound diagnosis is proven to be less invasive, even portable, at lower screening cost than most other medical imaging. However, being less visual than most, ultrasound image diagnosis is difficult even for trained professionals, and thus can benefit greatly from computer enhancement. Image texture analysis and clustering is improved from 8 days per image down to 5 seconds on a PC to enable point-of-care computer enhanced diagnosis as we- ll as more accurate tumor locating capabilities for treatment and biopsy guidance. Using this integrated approach, specific results for several sample patient image cases are tested and general conclusions are drawn, yet more images need to be tested. For future study, the color- coded cluster tissue sample images can potentially represent from normal to various degrees of seriousness of the abnormality in sample images similar to Gleason score for prostate cancer. Beside the current 5 Haralick´s textural features, wavelet and other faster image texture features are also under consideration for inclusion. The parameter settings of this efficient image texture analysis tool can also be tuned to cluster pixel groups in wide varieties of challenging micro/macro image applications.