Objective To develop and verify a deep learning algorithm for the automatic detection of anterior mediastinal lesions, aiming to improve the diagnostic efficiency of such lesions in chest CT examinations.
Methods A total of 256 patients who underwent chest CT examinations between 2015 and 2022 were enrolled from two medical centers, including 145 patients who presented anterior mediastinal lesions. Anterior mediastinal le sion contours were manually delineated by two senior radiologists. Three deep learning models were developed based on ResUnet with a multi-region-of-interest (MultiROI) strategy and data augmentation approaches for mediastinal lesions segmentation and false-positive reduction. The performance was evaluated with the DICE index, sensitivity, specificity, and free-response receiver operating characteristic (FROC) curves on both internal and ex ternal testing sets.
Results In the optimal model 3, the DICE scores were enhanced and reached 0.834 and 0.643 on the internal and external test sets, respectively. In the detection task, similar sensitivities of 0.794 and 0.773, and specificities of 0.893 and 0.836 were achieved on internal and external test sets, respectively. Of note, the false-positive rate was decreased to 0.125 and 0.101 per scan, with an average prediction time of 21.13-26.12 sec onds.
Conclusion The deep learning algorithm enables accurate segmentation and detection of anterior mediasti nal lesions on CT images, showing the potential to assist in the clinical diagnosis of anterior mediastinal lesions.
DownLoad: