Deep Convolutional Neural Network–based Software Improves Radiologist Detection of Malignant Lung Nodules on Chest Radiographs

Abstract

THORACIC IMAGINGC hest radiography is the most common radiologic examination, despite its inferiority to low-dose CT, for lung cancer screening (1). Some authors showed that up to 90% of "missed" lung cancer nodules can be found when the baseline chest radiograph is re-reviewed with the benefit of the follow-up examination showing the mass that has grown in size (2). Misdiagnoses of lung cancer can occur for many reasons. This oversight can be due to a lack of perception of the nodule, the decision to ignore a subtle density, and the satisfaction of search when another abnormality is identified (3-5). Lesion characteristics including size, density, and location make the detection of lung nodules more challenging on chest radiographs (6-8).To improve the efficacy of chest radiography for nodule detection, computer-aided detection (CAD) software has been developed and evaluated. In 2004, Kakeda et al (9) tested their CAD and reported that it was beneficial in analyzing radiographs with nodules but had an average falsepositive rate of 3.15 per image. de Hoop et al (10) showed Purpose: To compare the performance of radiologists in detecting malignant pulmonary nodules on chest radiographs when assisted by deep learning-based DCNN software with that of radiologists or DCNN software alone in a multicenter setting. Materials and Methods: Investigators at four medical centers retrospectively identified 600 lung cancer-containing chest radiographs and 200 normal chest radiographs. Each radiograph with a lung cancer had at least one malignant nodule confirmed by CT and pathologic examination. Twelve radiologists from the four centers independently analyzed the chest radiographs and marked regions of interest. Commercially available deep learning-based computer-aided detection software separately trained, tested, and validated with 19 330 radiographs was used to find suspicious nodules. The radiologists then reviewed the images with the assistance of DCNN software. The sensitivity and number of false-positive findings per image of DCNN software, radiologists alone, and radiologists with the use of DCNN software were analyzed by using logistic regression and Poisson regression.

Performance of a Deep Learning Algorithm Compared with Radiologic Interpretation for Lung Cancer Detection on Chest Radiographs in a Health Screening Population

“…In previous studies, the sensitivity of algorithms for lung cancer detection on chest radiographs increased when lesions were large and clearly demonstrated (12,16,18). Concordantly, our study also showed that the AUC increased when lung cancers were clearly visible on chest radiographs.…”

“…(a) they were tested using disease-enriched data sets (prevalence, 16%-75%), which were clearly unrealistic (16)(17)(18); (b) their test data sets were arbitrarily selected in terms of the size, number, and location of the lung cancers (16)(17)(18); and (c) their test data sets comprised clearly dichotomized cases (chest radiographs with lung cancer vs normal chest radiographs), which intentionally excluded any indeterminate chest radiographs or radiographs with other pathologies (16)(17)(18). By contrast, we performed our study in a real-world setting, using a real-world health screening population (23).…”

“…Recently, deep learning algorithms have staked out a place in lung cancer detection on chest radiographs (16)(17)(18) and have demonstrated excellent diagnostic performance in disease-enriched settings (16)(17)(18). Nam et al (16) reported that a deep learning algorithm achieved a sensitivity of 71%-91%, a specificity of 93%-100%, and an area under the receiver operating characteristic curve (AUC) of 0.92-0.99 in their validation data sets with a lung cancer prevalence of approximately 60%-68%.…”

“…Nam et al (16) reported that a deep learning algorithm achieved a sensitivity of 71%-91%, a specificity of 93%-100%, and an area under the receiver operating characteristic curve (AUC) of 0.92-0.99 in their validation data sets with a lung cancer prevalence of approximately 60%-68%. Sim et al (18) reported that…”

Challenges and solutions for introducing artificial intelligence (AI) in daily clinical workflow

“…Instead of being considered as stand alone actors, AI systems for diagnostic tasks in radiology should thus be considered as a complement to the radiologist. Studies evaluating these systems should concentrate on measuring and monitoring how the use of AI enhances the performance of radiologists [4].…”