Computer aided diagnosis of Lymphoma based on dual-mode ultrasound radiomics

Objective To evaluate the diagnostic performance of dual-mode ultrasound radiomics in differential diagnosis of lymphoma based on computer aided elastography and B-mode images. Method The Bmode ultrasound and elastography images from 543 lymph nodes (142 benign, 258 lymphoma and 143 metastatic) in 538 patients were analyzed retrospectively. Radiomic features were extracted from dual-mode images by computer. These features were selected by selection methods based on information theory. Different lymphadenopathy subsets were classified by the support vector machine. Then different modalities and different radiomic feature subsets were amalgamated Adaboost algorithm, and resulted in the improvement of lymph node classification. Results Radiomics features was selected to classify different lymphadenopathy subsets, including four B-mode ultrasound (minor axis; long/minor axis rate; homogeneity; solidity) and three elastography features(hard area ratio; strain rate; coefficient of variation, Cov). The classification of lymphoma subset from other subsets were performed statistically significant by both homogeneity(P 0.05). Moreover, the area under the receiver operating characteristic curve of multi-class classification counted by Adaboost algorithm were 0.875 in the discrimination of lymphoma subset from benign subset and 0.843 in the discrimination of lymphoma subset from metastatic subset, respectively. Conclusion Radiomics features derived from dual-mode ultrasound images were valuable for diagnosis of lymphoma. Especially, the homogeneity in B-mode ultrasound and the Cov in elastography which doctors could not identify by unaided eyes were clearly analyzed by computer.