Towards high performance cell segmentation in multispectral fine needle aspiration cytology of thyroid lesions

Thyroid nodule is a common cancer of the thyroid gland that affects up to 20% of the world population and approximately 50% of 60-year-old persons. Early detection and screening of the disease, especially analysis by fine needle aspiration cytology (FNAC), has led to improved diagnosis and management of the disease. Simultaneously, advances in imaging technology has enabled the rapid digitization of large volumes of FNAC specimen leading to increased interest in computer assisted diagnosis (CAD). This has led to development of a variety of algorithms for automated analysis of FNAC images, but due to the large scale memory and computing resource requirements, has had limited success in clinical use. In this paper, we present our experiences with two parallel versions of a code used for texture-based segmentation of thyroid FNAC images, a critical first step in realizing a fully automated CAD solution. An MPI version of the code is developed to exploit distributed memory compute resources such as PC clusters. An OpenMP version is developed for the currently emerging multi-core CPU architectures, which allow for parallel execution on every desktop system. Experiments are performed with image sizes ranging from 1024 x 1024 pixels up to 12288 x 12288 pixels with 21 spectral channels. Both versions are evaluated for performance and scalability.