Development and validation of a self-attention network-based algorithm to detect mediastinal lesions on computed tomography images.

Background: Diagnosis of mediastinal lesions on computed tomography (CT) images is challenging for radiologists, as numerous conditions can present as mass-like lesions at this site. This study aimed to develop a self-attention network-based algorithm to detect mediastinal lesions on CT images and to evaluate its efficacy in lesion detection. Methods: In this study, two separate large-scale open datasets [National Institutes of Health (NIH) DeepLesion and Medical Image Computing and Computer Assisted Intervention (MICCAI) 2022 Mediastinal Lesion Analysis (MELA) Challenge] were collected to develop a self-attention network-based algorithm for mediastinal lesion detection. We enrolled 921 abnormal CT images from the NIH DeepLesion dataset into the pretraining stage and 880 abnormal CT images from the MELA Challenge dataset into the model training and validation stages in a ratio of 8:2 at the patient level. The average precision (AP) and confidence score on lesion detection were evaluated in the validation set. Sensitivity to lesion detection was compared between the faster region-based convolutional neural network (R-CNN) model and the proposed model. Results: The proposed model achieved an 89.3% AP score in mediastinal lesion detection and could identify comparably large lesions with a high confidence score >0.8. Moreover, the proposed model achieved a performance boost of almost 2% in the competition performance metric (CPM) compared to the faster R-CNN model. In addition, the proposed model can ensure an outstanding sensitivity with a relatively low false-positive rate by setting appropriate threshold values. Conclusions: The proposed model showed excellent performance in detecting mediastinal lesions on CT. Thus, it can drastically reduce radiologists' workload, improve their performance, and speed up the reporting time in everyday clinical practice.