Auxiliary tumour diagnosis image with deep learning technology

• Published in: The Journal of supercomputing Vol. 78; no. 1; pp. 578 - 595
• Main Author: Hou, Shengchao
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.01.2022; Springer Nature B.V
• Subjects: Accuracy; Algorithms; Back propagation networks; Compilers; Computed tomography; Computer Science; Deep Learning; Density; Diagnosis; Genetic algorithms; Image quality; Image reconstruction; Interpreters; Lung cancer; Machine learning; Medical imaging; Metastasis; Neural networks; Parallel Computing in Biomed Sciences & Healthcare; Processor Architectures; Programming Languages; Sensitivity; Tumors; Malignant tumour of lung cancer; Primary lesion metastasis; Depression; CT imaging data; U-Net neural network
• ISSN: 0920-8542; 1573-0484
• DOI: 10.1007/s11227-021-03881-7

In order to analyse the application value of U-Net neural network in reconstruction and diagnosis of computed tomography (CT) scanning image of lung cancer and provide assistance for CT image diagnosis of lung cancer, a new deep learning-based tumour image diagnosis algorithm U-Net-NT (U-Net-Newton’s method) was constructed based on the U-Net neural network framework and compared with genetic algorithm-back propagation (GA-BP), random forest (RF), Semi-Naive Bayesian (SNB), and filtered back projection (FBP). In addition, the information given by experts was taken as a control, set as D0 group. Tumour disappearance rate (TDR), maximum standard uptake value (SUVmax), and metabolic tumour volume (MTV) were calculated with True D. The results showed that the diagnosis results on primary lesion density, SUVmax, and MTV level in the GA-BP group, RF group, SNB group, and FBP group were significantly lower than that in the D0 group ( P  < 0.05); the diagnosis results of primary lesion density, SUVmax, and MTV level in the U-Net-NT group had no obvious difference with the D0 group ( P  > 0.05); the U-Net-NT group had a significantly higher Dice coefficient than the GA-BP group, the RF group, SNB group, and FBP group, and the differences were statistically significant ( P  < 0.05); the volumetric overlap error (VOE) and running time in the U-Net-NT group were observably lower than those in the GA-BP group, RF group, SNB group and FBP group, and the differences were statistically significant ( P  < 0.05); the accuracy, sensitivity, and specificity for the diagnosis of primary lesion lymph node metastasis, pleural metastasis, and bone metastasis in the U-Net-NT group were not significantly different from those in the GA-BP group, RF group, and SNB group, FBP group ( P  > 0.05), and the accuracy, sensitivity, and specificity for the diagnosis of primary lesion brain metastasis and liver metastasis were higher than those in the GA-BP group, RF group, SNB group, and FBP group significantly ( P  < 0.05). In short, the new deep learning-based U-Net-NT algorithm constructed in this study is more excellent in the accuracy, sensitivity, and specificity for the diagnosis of the primary lesion level and metastasis and has better quality of reconstructed image.