Improved image segmentation for tumor volume quantification

Abstract

Tumor segmentation and quantification are part of the main issues in Image Guided Surgery (IGS) / Computer Aided Surgery (CAS) procedure relating to brain tumor ablation. Lack of accurate segmentation has often led to either insufficient tumor resection or over-ablation resulting in tumor recurrence or damage to healthy cells. Other problems associated with IGS are inadequate knowledge of tumor volume, long perioperative time, and high cost. This research work addressed these issues in the following ways. In respect of the first issue, earlier researchers have shown the inadequacy of a single segmentation method in addressing medical image segmentation problems, which invariably affects subsequent analysis of the image concerned. Furthermore, their works suggest the possibility of an accurate and fast image segmentation technique via hybridization of various existing methods. This is the approach taken by this research which intelligently combines multiplethresholding, template matching involving 2D correlation, and active contour using level-set algorithm (LSA) segmentation methods to derive a hybrid and multi-level segmentation technique called Improved Image Segmentation Technique (IIST). IIST is 96.6% accurate compared with 'ground truth', faster, and semi-automatic. Regarding inadequate knowledge of tumor volume, the approach used in this aspect of the research is the combination of clinical information about image slices with known mathematical approaches. This led to methodological formulation of tumor quantification method (TQM) that accurately determines the volume of brain tumor in patients' image record. In the said approach, individual tumor area (TA) - areas of regions of interest from each of the patient's tumor slices surface - was tabulated with the respective depth of the slice called slice thickness (ST). Then, this was use to accurately (96.7%) determine tumor volume. These two techniques were then fused to form the bi-level segmentation and tumor quantification scheme that works by firstly transforming grey-scale (GS) image into binary-space (BS) image. Then with the aid of a modified form of level-set algorithm called (MLSA), the scheme accurately encircles the region of interest (ROI) - Tumor. Finally, the mathematical algorithms and functions developed were implemented on a parallel computing network. The use of parallel computing platform further enhanced computation time which reduces the perioperative time and secondly, provided a cheaper means of acquiring IGS infrastructures; thereby reducing the cost of implementing IGS. The results obtained were evaluated on three cardinal points namely: algorithm simplicity, segmentation time, and segmentation accuracy. The evaluation results show that MLSA on BS image outperforms LSA on GS image in terms of segmentation time, and segmentation accuracy. More so, MLSA is less complex compared to LSA because of the elimination (in MLSA) of the factor responsible for the mean intensity of pixels exterior to the zero level-set curves. In addition to the contribution of this research work which are; simpler algorithm, fast, accurate, and semi-automatic tumor segmentation and quantification scheme, the whole scheme itself is a novel contribution as this is the first time that this approach consisting of the developed discrete integral function, modified level-set algorithm, and graphical method are used in tumor quantification.