Clustering of Medical Images for Analysis: A Fuzzy Approach

• DOI: 10.5339/qfarc.2016.hbpp2825
• Publication Date: 2019-07-31T07:52:45Z
• Authors: Sarada Prasad Dakua; Nikhil Navkar; Julien Abinahed; Georges Younes; Abdulla A. Al-Ansari
• Keywords: Feature (linguistics)

Background and Objective Often times, clinicians use a three-dimensional set of medical images to diagnose plan treatments, which typically include visual identification structures such as bones tissues [1]. This can be challenging task anatomical interest contain significant noise, easily blend with neighboring tissues. We propose tackle 2 cases: (a) treatment planning pelvic fractures where small size ring formed by the fused ischium, ilium, pubis attached sacrum contains vital (including major blood vessels, nerves, digestive reproductive organs) should carefully delineated, (b) liver cancer malignant tissue has removed. The pixel intensity tumor is similar those healthy proper delineation utmost important before proceeding therapy. To address aforementioned challenges, in this work we present soft-clustering technique using Enhanced Fuzzy C-Means (EFCM) along bilateral filter detect region interest. key feature proposed algorithm combines domain range filtering allowing maintain balance between preservation relevant details degree noise reduction. approach allows traditional not only exploit useful spatial information, but also dynamically minimize clustering errors caused common images. Methodology A three-step workflow used process images: Step 1: After MR/CT are acquired; clinician initially draws rough outline around (where fracture or present) on two-dimensional image slices. manual input reduces computational time determine desired cluster providing instead scanning/processing entire image. 2: In step, remove while preserving edges [2]. Linear filters, Gaussian, compute weighted average values neighborhood. weights decrease distance from neighborhood center. works well for local pixels have (slow variation). As that corrupts these mutually less correlated than signal, averaged away signal preserved. However, assumption slow variations fails at edges, consequently blurred linear low-pass filtering. context, non-linear/bilateral both smooth regions, within each other, normalized similarity function close one. Consequently, acts essentially standard filter, averages weakly differences preserves edge details. 3: last EFCM applied noise-filtered assigning membership data point respect centers [3]. computed center point. towards particular varies linearly per distance. Closer center, higher its membership. summation across different clusters equal An objective based Euclidean metric then update iteratively. parameter estimation resulting described may robust noisy environment. Therefore work, develop an uses modified term (described Table I) against meaningful compact analysis (Fig. 1). Results Conclusion method was evaluated two datasets: CT (two subjects) publicly available online research purposes, OsiriX website ( http://www.osirix-viewer.com/datasets ). acquisition followed: Slice Thickness: mm, Pixel Spacing: 0.29 mm × Bit-depth: 12, Acquisition Matrix: 512 512. containing five anonymized subjects, obtained Hamad Medical Corporation, Doha, Qatar. 3 0.32 16, algorithms were implemented MATLAB R2013a running workstation 16 GB RAM 2.8 GHz Intel processor. required perform segmentation recorded 8 ± 1.5 minutes. However could further reduced, implementation done without optimization internal calls. initial assessment experimental results shown satisfactory outcomes cases filters datasets failed identify values. modification lead soft identifying future, optimize validate extensively tissue-types multiple imaging modalities. References [1] Vona G. et al., “Impact cyto-morphological detection circulating cells patients cancer,” Hepatology, 39, 792–797, 2004. [2] Sugimoto K. “Compressive Bilateral Filtering,” IEEE Trans. Image Processing, 24, 3357–3369, 2015. [3] Havens T. C., “Fuzzy c-Means Algorithms Very Large Data,” Systems, 20, 1130–1146, 2012.