Session Item

A major health concern for subjects with lung cancer is the decrease in bone marrow and blood cells during and after thoracic radio(chemo)therapy. Segmentation is an important and challenging task in the automatic image analysis of bone marrow. This project addresses the design and preliminary analysis of an automatic bone marrow contouring algorithm. 

For the development of the automated algorithm, a technique of statistical methods and Max-flow has been used. This technique combines gray information and statistical information derived from combined distance histograms to feed a convex relaxation algorithm for fast and accurate segmentation of bone structures. Two different segmentation methods have been carried out that differ in the initial treatment of the image: 1) The first is based on performing thresholding. The objective of thresholding is to separate the set of data available in the image into two levels or main values, so that they are differentiated from the rest of the values ​​present. A value of 60UH has been used for the threshold, so that all values ​​below this are discarded, except those that are within pixels of higher values. In this way, only the bone structures (high UH values) and the bone marrow (pixels included within the bone contours) are highlighted. To differentiate between bone and marrow, a new thresholding is performed with a value of 315 UH that separates the bone marrow and bone pixels below and above said value respectively; 2) The second method works directly with the available values ​​from the image and calculates a mean value for bone marrow and bone. Extreme values ​​below 60UH and above 1500UH are ruled out.

We first segmented manually 2 cases and were used for training and evaluation purposes. The automated segmentation of the bone marrow with Method 1 conforms more precisely to the contours made manually than automated segmentation with Method 2, where the structures marked as bone marrow exceed the manual contours to a greater extent and are located in areas in which there is no bone marrow such as the mediastinal region and cartilage sacks. This is shown in the Table with significantly higher sensitivity and positive predictive values ​​for Method 1 (0.39 and 0.55 respectively in Case 1; 0.55 and 0.66 for Case 2) compare with Method 2 (0.03 and 0.21 in Case 1; 0.12 and 0.47 in Case 2). The DICE and Jaccard coefficients are also of higher value, indicative of a greater similarity between the segmentations of the two cases performed with Method 1 than those performed with Method 2.

The preliminary results of the current work regarding the development of the automated algorithm for bone marrow segmentation, although limited, provide a promising starting point from which to develop a robust segmentation algorithmThe algorithm will help physicians and physicists in the radiation planning and the analysis of potential constraints associated with haematological toxicity.