Hybrid CPU and GPU computation to detect lung nodule in computed tomography images

Lung Nodule is a white patch on the thorax medical image, usually used as an early marker of lung cancer. This research aims to produce algorithms that can detect lung nodules automatically in CT images, by utilizing a combination of hybrid computing between Central Processing Unit (CPU) and Graphical Processing Unit (GPU). The framework used is Compute Unified Device Architecture, which consists of platform and programming model. The algorithm consists of several steps; read DICOM and data normalization, lung segmentation, candidate nodule extraction, and classification. Normalization is required to facilitate calculation by changing the data type ui16 to ui8. Furthermore, segmentation is used to separate the lung parts with other organs, where at this stage the Otsu Algorithm and Moore Neighborhood Tracing (MNT) are used. The next step is Lung Nodule Extraction, which aims to find the nodule candidate. The last step is a classification that utilizes the Support Vector Machine (SVM) to distinguish which one is nodule or not. The algorithm successfully detects near round nodules that are free-standing or not attached to other parts of organs. After undergoing ground truth tests, it was found that under some conditions, the algorithm has not been able to distinguish nodules and other strokes that resemble nodules. While in terms of computing speed is found a very surprising result because overall single CPU computing provides better results compared to hybrid CPU and GPU computing. Multiple morphology and transmission time to GPU contributed to the double execution time of hybrid model compared to single CPU. Adjustment in dataset grouping by detecting the nodule simultaneously for several dataset will also improve the performance of hybrid CPU and GPU computation.