Development and validation of a deep learning algorithm detecting 10 common abnormalities on chest radiographs

Abstract

We aimed to develop a deep learning algorithm detecting 10 common abnormalities (DLAD-10) on chest radiographs, and to evaluate its impact in diagnostic accuracy, timeliness of reporting and workflow efficacy.DLAD-10 was trained with 146 717 radiographs from 108 053 patients using a ResNet34-based neural network with lesion-specific channels for 10 common radiological abnormalities (pneumothorax, mediastinal widening, pneumoperitoneum, nodule/mass, consolidation, pleural effusion, linear atelectasis, fibrosis, calcification and cardiomegaly). For external validation, the performance of DLAD-10 on a same-day computed tomography (CT)-confirmed dataset (normal:abnormal 53:147) and an open-source dataset (PadChest; normal:abnormal 339:334) was compared with that of three radiologists. Separate simulated reading tests were conducted on another dataset adjusted to real-world disease prevalence in the emergency department, consisting of four critical, 52 urgent and 146 nonurgent cases. Six radiologists participated in the simulated reading sessions with and without DLAD-10.DLAD-10 exhibited area under the receiver operating characteristic curve values of 0.895-1.00 in the CT-confirmed dataset and 0.913-0.997 in the PadChest dataset. DLAD-10 correctly classified significantly more critical abnormalities (95.0% (57/60)) than pooled radiologists (84.4% (152/180); p=0.01). In simulated reading tests for emergency department patients, pooled readers detected significantly more critical (70.8% (17/24) versus 29.2% (7/24); p=0.006) and urgent (82.7% (258/312) versus 78.2% (244/312); p=0.04) abnormalities when aided by DLAD-10. DLAD-10 assistance shortened the mean±sd time-to-report critical and urgent radiographs (640.5±466.3 versus 3371.0±1352.5 s and 1840.3±1141.1 versus 2127.1±1468.2 s, respectively; all p<0.01) and reduced the mean±sd interpretation time (20.5±22.8 versus 23.5±23.7 s; p<0.001).DLAD-10 showed excellent performance, improving radiologists' performance and shortening the reporting time for critical and urgent cases.

FIGURE 1

Development and validation of DLAD-10. SNUH: Seoul National University Hospital; ILD: interstitial lung disease; CT: computed tomography. See main text and supplementary figure E1 for details of the training stage.

FIGURE 2

Examples of DLAD-10 output. a) Each of 10 possible abnormalities was localised and displayed with its probability score. Urgency categorisation was performed based on the most urgent abnormality. This image was categorised as critical as it contained pneumothorax (Ptx) (in addition to nodule (Ndl) and pleural effusion (PEf)). b) A 47-year-old female patient visited the emergency department complaining of vague chest pain. A small pneumoperitoneum (Ppm) was detected by DLAD-10, while no readers detected the lesion in the conventional reading session. In the DLAD-10-aided reading session, all readers detected pneumoperitoneum. c) A 24-year-old male patient visited the emergency department due to left chest pain. A small left pneumothorax (Ptx) was detected by DLAD-10. Three readers reported pneumothorax in the conventional reading session and all six readers reported it in the DLAD-10-aided reading session. The arrows on the computed tomography scans in b) and c) indicate the corresponding abnormalities visualised on the chest radiographs.

FIGURE 3

Results of DLAD-10 and three thoracic radiologists for the Seoul National University Hospital external validation dataset: the area under the receiver operating characteristic curve (AUROC) of DLAD-10 and the performance of each radiologist are presented for each abnormality. a) Pneumothorax, b) pneumoperitoneum, c) mediastinal widening, d) nodule, e) consolidation, f) pleural effusion, g) atelectasis or fibrosis, h) calcification and i) cardiomegaly.