Auxílio computadorizado ao diagnóstico do câncer de pulmão otimizado por GPU

Abstract

Lung cancer is the leading cause of cancer-related deaths in the world and its main manifestation occurs due to the appearance of pulmonary nodules. Pulmonary nodule classification is a complex task because it is a subjective and qualitative process that might be compromised by interpretation errors. Therefore, it is important to integrate computational tools to the pulmonary nodule classification process, since they have the potential to characterize objectively and quantitatively the lesions, so they can aid lung cancer diagnosis process. Content-Based Image Retrieval (CBIR) has been described as one of the most promissing differential diagnosis tool, since it is capable of retrieving similar cases from large image databases that were previously diagnosed. However, CBIR has some limitations, like the image feature extraction process and the time to compare one reference image with an image database. In this context, the goal of this work is to develop an algorithm to aid the diagnosis of lung cancer and the pulmonary nodule classification, using CBIR with the integration of three methods: 3D Texture Analysis and 3D Margin Sharpness Analysis for nodule characterization, and optimization on the execution time of the nodule comparison with paralelism on a Graphics Processing Unit (GPU). Images used in this work were computed tomography scans provided by the Lung Image Database Consortium, which has pulmonary nodules identified and classified by specialists according to the lesion’s likelihood of malignancy. Texture Attributes (TA) were extracted from a co-occurrence matrix obtained from the nodule volume. Margin Sharpness Attributes (MSA) were extracted from perpendicular lines drawn over the borders on all nodule slices. Integrated Attributes (IA) were created by concatenating TA and MSA. Euclidean distance was employed as similarity metric between feature vectors. CBIR algorithm’s precision was evaluated on the 10 most similar cases according to the likelihood of malignancy and with Precision and Recall parameters, on single-core, multi-core and many-core architectures. Results showed that MSA obtained more efficiency on pulmonary nodule retrieval, in the majority of precision evaluation scenarios, with the precision increase of 2% compared with TA. Texture attributes obtained the same efficiency as IA and presented higher mean precision only on benign nodule retrieval with Precision vs. Recall metric, with the precision increase of 3% compared with MSA. Results also showed that GPU, represented by the many-core device, was able to decrease execution time on image feature vector comparison and increase Euclidean distance performance on pulmonary nodule retrieval, with speedups of 16x, 17x and 19x. Therefore, CBIR allied to 3D margin sharpness descriptors and GPU optimization have big potential as a computer-based tool on lung cancer diagnosis and pulmonary nodule classification.