Construction of differential diagnosis and staging model of pneumoconiosis based on multi-task learning.

Objective To construct a deep learning model based on multi -task learning to assist clinicians in differential diagnosis and staging of pneumoconiosis. Methods The digital chest radiographs of 3 600 patients from an occupational disease hospital in Sichuan Province from 2011 to 2022 were collected, and the full convolution neural network (UNet) was used to segment the lung field. Based on multi-task learning, the multi-task model was constructed using the correlation between tasks. The multi -task model was pre -trained on the ChestX -ray14 dataset, whose backbone network was DenseNetl21, and two classifiers were added behind the backbone network. Paired t-test was used to compare the differences in accuracy, precision, sensitivity, and Fl scores between single-task model and multi-task model. Results The test set results showed that the differential diagnosis and diagnostic staging performance of the single-task model was about 90% and 77%, respectively. The differential diagnosis and diagnosis staging performance of the multi-task model was about 94% and 86%, which was higher than that of the single-task model about 4% and 9%, respectively. The difference between the evaluation indexes was statistically significant (P < 0.05). Conclusion The multi-task model has more advantages than the single-task model and can effectively realize the differential diagnosis and accurate staging of pneumoconiosis and pulmonary tuberculosis. [ABSTRACT FROM AUTHOR]