Under both research and clinical settings dealing with measurements and analysis of tumors using diffusion weighted imaging (DWI), radiologist input for manual demarcation of tumors on each scan is usually not feasible. The process is cumbersome and not time effective. The purpose of this study was to identify and implement an algorithm that can be used to segment a tumor from a given stack of diffusion weighted MR images, optimizing on speed and accuracy with minimal manual input requirements.DWI dataset from six patients (Age=25.3±8.8ys) with Ewing’s Sarcoma were acquired under the Institutional Review Board approved protocol. The acquisition was performed using 1.5T Philips Achieva MRI scanner with Spin Echo Planar imaging (SP-EPI) sequence with TE=66msec,TR= 1782msec, 5mm slice thickness and 144x144 matrix size. Images with b=800 s/mm2 were used for segmentation analysis. The segmentation algorithms and analysis were implemented in an in-house built toolbox in MATLAB. Three segmentation algorithms were examined in this study: 1) Otsu thresholding [1], 2) Otsu threshold-based region growing [2] and 3) Active contours using Chan-Vese implementation [3]. With these three base segmentation algorithms further three methodologies were evaluated for the accuracy and computational processing time: i) Fully Automated Segmentation, ii) Semi-automated Segmentation and iii) Manual Segmentation.

i) Fully automated segmentation, requiring no human input. This algorithm performs the base technique on each slice of the MRI image stack, without using any additional human inputs. It is typically expected to save time and manual effort.

ii) Semi-automated, requiring human input on selected slices with complex anatomy or structural pathology. This algorithm optimizes between time, accuracy and manual effort, and it is implemented as follows: a) The user is required to input a number of seed points (in case of Otsu-threshold based region growing) or a tumor Region of Interest (ROI) (in case of Active Contours based region growing) on the first slice of the MRI image stack. b) The algorithm uses intelligence to initialize the subsequent slices with prospective seed points or Tumour ROI. It calculates the centroid of ROI in each slice, maintains the continuity to the subsequent slices covering complete 3D anatomy. The size of the segmented tumour in each slice is calculated in proportion to the whole anatomy in the slice. A massive change in tumour size in adjacent slices (>50%) is considered to be abnormal and a request is made for manual confirmation to either correct the ROI or agree to proceed for further segmentation.

iii) Manual, requiring human input on each slice. This algorithm typically optimizes accuracy over time and manual effort.

Tumour ROIs drawn manually by a radiologist (clinical experience of > 5 years in cancer imaging) were considered as standard tumour mask and used to calculate the error of the different segmentation algorithms. The accuracy measure used to compare the performance of these algorithms based on misclassification was calculated as follows: $$$Error=1-(\frac{A\cap B}{A\cup B})$$$, where A is the area demarcated as the tumor by the expert radiologist, and B is the area demarcated by the segmentation algorithm as tumor.

Figure 1 presents a performance summary of segmentation algorithms, with the average performance error across the 3D image stacks and the execution time per slice. The automated Otsu thresholding, while being instantaneous and requiring no manual input, delivered a high 70% error. Manual and semi-automated Otsu threshold based region growing algorithm showed a moderately accurate result, with average error of 30-37%, which was considerably better than the fully automated algorithm. It typically consumes ~1-2 second per image slice. Manual and semi-automated Active contours based region growing algorithm delivered more accurate result than the other two methods with error rate of ~27-29%. It typically takes ~4-5 seconds per slice. Figure 2a shows one DW image at b=800s/mm2 from one of the representative data set and Figure 2b presents image mask drawn for tumour by the radiologist. An illustrative example of the segmented images corresponding to the original image (Fig. 2a) for each algorithm is depicted in Figure 2c-g. Both Semi-automated and manual algorithms have qualitatively more accurate segmentation than automated method. Manual active contour segmentation has the least error in segmentation but is computationally expensive and requires expert intervention at multiple stages. Semi-automated algorithms based on both Otsu-threshold based and Active Contours based region growing methods had reasonably accuracy (<30% error) and required much less manual intervention. Thus semi-automated segmentation methods may be more useful for tumour demarcation on DW images with acceptable level of accuracy.