Stromov G., Postgraduate Student at Medical and Industrial Department, NI TPU, Tomsk

Ryzhkov D., Postgraduate Student at Medical and Industrial Department, NI TPU, Tomsk

Fokin V., Doctor of Sciences, Professor at Medical and Biological Cybernetics Department, SSMU, Tomsk

Evtushenko G., Doctor of Sciences, Head of Medical and Industrial Department, Professor at NI TPU, Tomsk

Criteria for Optimal Regions of Interest Retrieval on 3D Medical Images using Integral Assessment Method

The common approach for automated ROI retrieval in 3D MRI images via Integral Assessment (IA) [2] Method was shown in [3]. Model MRIs provided by BrainWeb [1] resources were used as source data. These MRI are generated based on two phantoms: the normal one and with a pathology manifestation (severe multiple sclerosis). Histograms of assessment distribution show that segments contained the morphological substratum have greater values comparing to ones without a manifestation. This could be used as a basis for ROI retrieving in studied images. In this publication we will show a reconstruction method of the morphological substratum and optimal criteria to do this depending of noise level on data.

Source data are monochrome 8bit or 12bit 3D images with 181×217 pixels in the traversal section and with an adjustable step from 1 to 10 pixels on the axial direction (181 pixels as a limit). The lesions are clearly observed in any scan mode, we have chosen T1-weighted images. In this mode lesions are presented as hypointensive regions (refer to Figure 1). Matrices for calculations are formed by the following way: source files are split on small equal segments (flat or volumetric ones, or voxels) with equal seeds from the beginning of coordinates for each one, and then these block are unwrapping into a vector of normalized values of pixels' brightness. It is important to note that a way of unwrapping doesn’t affect IA values, the principal moment is in an uniformity of the unwrapping procedure for all the images. Obtained by such way vectors are aggregated into 2d arrays presented the referent or estimated state (depending on the phantom), respectively.

Series of calculations were performed for a certain noise level (3%, 5%, 7%, 10%, 11%, 14% and 15%). A character of the IA distribution was analyzed for each one. We've found out that there are a few areas in sensitivity depending on the noise level:

· High sensitivity area: with level less or equal to 9% the morphological substratum is detected in all slices and with any size of blocks;

· Conditional sensitivity area: ROIs are showed in results, but not in all slices and/or not with all block sizes with noise level from 10% to 14%;

· “Blind” area: there are no ROIs in result starting from 15% of noise.

Optimality of partitioning size depends on the goal of study. If we need to get a detailed description of ROI, it will be better to chose small size of blocks. In this case the heterogeneity of studied region will necessarily be shown as adjacent portions with high but different brightness on the IA heatmap (Figure 2).

We can reduce the number of artifacts caused by noise by increasing partitioning size and save an acceptable level of detail (Figure 3). The high level of noise makes it impossible to get detailed information about ROIs, and using the “big” blocks is the sole way to detect a manifestation of pathology (high level of noise is also could be treated as small differences in the compared states).

The study revealed correlations between noise level and detailed elaboration of ROI in the result images. If there is information about adjacency between detected ROIs we can build 3D reconstruction of a morphological substratum (Figure 4).

References:

1. BrainWeb Simulated Database. URL: http://brainweb.bic.mni.mcgill.ca/brainweb/ (last accessed at 01/20/2014).

2. Fokin V.A. Tehnologija integral'noj ocenki sostojanija biomedicinskih sistem // Sistemy upravlenija i informacionnye tehnologii, 2008. – V(31). – P. 191-194.

3. Stromov G. G., Fokin V. A., Evtushenko G.S. Integral'naja ocenka trehmernyh biomedicinskih izobrazhenij s ispol'zovaniem tehnologii raspredelennyh vychislenij // Biotehnosfera, 2012. – V 3-4. – P. 68-72.

Figure 1. The normal (the top of the image) and pathological (the bottom of the image) phantoms are shown in sagittal (A), transversal (B) and axial (C) sections. A morphological substratum is pointed by arrows.

Figure 2. Heatmap (A) of the IA distribution for blocks sized by 2×2×2 and 3% noise. (B) shows the same heatmap after 99% of IA values were cut. In (C) these blocks were painted white for better visibility.

Figure 3. ROIs are obtained for blocks sized 4×4×3 and 5% noise.

Figure 4. An isosurface reconstruction for a morphological substratum obtained for blocks sized 2×2×2 and 3% noise.