| Resumo: | Cancer is a leading cause of death worldwide. Treatment strategies rely on accurate tumor staging and surveillance by imaging screening. Whole-body Diffusion Weighted Imaging (DWI) has high value to detect, characterize and quantify malignancies with irregular diffusion patterns, such as Multiple Myeloma (MM). However, the large volume of imaging data hinders the reading process. Manual delineation (segmentation) of tumor sites becomes a time-consuming process and lacks reproducibility. The lack of adequate tools in clinical practice leads radiologists to perform only a qualitative description of DWI images and measure of the biggest lesion diameter, an inherently subjective process. Arising from this need, this dissertation aimed to develop an algorithm to improve the process of segmentation of lesions ofMMDWI images, to allow accurately and rapidly tumor burden quantification, by validating against radiologist’s manual segmentation. Quantification of bone lesions (hyperintense on DWI) volume without considering normally hyperintense organs was made possible due to the development of an atlas-based and a smart lesion detector algorithm. The first allowed the removal of normal hyperintense organs from the images to be studied, using a suitable registration procedure. The second applied an outlier detector algorithm and compared voxel-by-voxel and connected-component approaches on different b-value images (directly acquired and computed), to delineate lesions. T1-weighted images were also used to improve lesion detection. The atlas-based algorithm revealed good alignments against the manual segmentation: Dice Similarity Coefficient (DSC) of 0.63 0.03 for male and 0.58 0.05 for female. Regarding lesion detection, the connected-component approach applied to the directly acquired b-value image was the method that presented the greatest similarity to the gold standard. Although not yet overcoming the manual segmentation performance, these results are suggestive of the great potential of semi-automatic registration methods combined with quantitative algorithms to analyze DWI images, assisting radiologists while defining tumor burden. Staging, prognosis and response analysis in several pathologies may be facilitated.
|
|---|