• artificial intelligence
• chest x-ray
• deep learning
• education

Chest X-rays (CXRs) are fundamental in the diagnosis of various thoracic pathologies, offering crucial insights into lung conditions. Yet, accurately and efficiently interpreting CXRs poses a challenge due to the intricate nature of thoracic anatomy, the overlapping structures, and the diverse array of potential abnormalities. Integration of AI-powered deep learning software in CXR interpretation holds promise as a solution, augmenting diagnostic precision and expediting the detection of intricate multiclass findings. Additionally, these AI tools serve as invaluable educational resources for radiology residents, fostering their growth and competence in interpreting thoracic CXRs. This study aimed to evaluate the impact of AI-based tools on diagnostic performance and confidence of radiology residents in detecting multiclass findings on CXR.

In our institution, a retrospective review was conducted on all CXRs performed in June 2023 at the 2nd Radiology Unit in Pisa University Hospital, using qXR software (Qure.ai®) for labeling. The software output comprised the original postero-anterior radiographic image overlaid with segmentation and labels indicating the identified findings. A radiologist with over a decade of expertise in chest radiography was considered as the gold standard for this study.

Each CXR was evaluated by two radiology residents, one in the first year and the other in the fourth year of their residency training, for the presence of ten different findings, including blunted costophrenic angle, cardiomegaly, cavity, consolidations, fibrosis, hilar enlargement, nodules, opacities, pneumothorax, and pleural effusion. The residents rated their confidence levels on a scale from 0 (completely uncertain) to 10 (absolutely certain). Following this, the processed image from qXR was provided, enabling the radiology trainees to re-evaluate based on this input and restate their confidence levels. Parameters such as accuracy, sensitivity, specificity, positive predictive value, and negative predictive value were calculated for qXR and for each radiology resident, both with and without qXR.

Statistical analyses were conducted using SPSS v.28.0, employing the t-student paired test to compare the diagnostic confidence levels before and after the adoption of qXR and the t-student unpaired test to compare diagnostic confidence levels between the first-year and fourth-year radiology residents. A p-value of less than 0.05 (two-tailed) was considered statistically significant.

qXR exhibited outstanding performance in terms of accuracy, sensitivity, specificity, positive predictive value, and negative predictive value. Both radiology residents showed improvement in diagnostic accuracy and diagnostic confidence levels using qXR. Specifically, a statistically significant difference (p<0.005) was observed in the level of diagnostic confidence before and after the adoption of qXR (Qure.ai®) for both radiology residents. Moreover, the level of diagnostic confidence was significantly higher (p<0.05) for the fourth-year radiology resident compared to the first-year radiology resident.

In conclusion, qXR has showcased its potential role within the residents' training program, affirming its valuable contribution to enhancing diagnostic accuracy and confidence in the interpretation of multiclass findings on CXR.