Diagnostic Value of Artificial Intelligence in Early-Stage Lung Cancer
Lung cancer remains the leading cause of cancer-related morbidity and mortality worldwide. Early diagnosis and treatment are critical to improving the 5-year survival rate of patients with lung cancer. Artificial intelligence (AI), particularly deep learning, has emerged as a promising tool in medical imaging, offering efficient self-optimization capabilities. AI not only enhances the recognition of pulmonary nodules but also improves the diagnostic efficiency of early-stage lung cancer. This study evaluates the diagnostic value of AI in early-stage lung cancer, comparing its performance with that of radiologists and exploring the potential benefits of combining AI with human expertise.
The study focused on patients who underwent pulmonary nodule surgery at Zhongshan Hospital, Fudan University, between January 2016 and December 2018. The inclusion criteria were stringent: (1) a definitive surgical pathological diagnosis, (2) clear and qualified thin-layer chest computed tomography (CT) imaging data within one week before the operation, (3) solitary pulmonary nodules (SPN) with a diameter between 5 mm and 30 mm, and (4) pathological staging of malignant nodules showing tumor in situ (Tis) or stage IA. Based on clinical experience, the sensitivity and specificity were estimated to be 65%. A total of 360 nodules, comprising 180 malignant and 180 benign cases, were selected. To balance confounding factors such as gender and age, propensity score matching (PSM) was conducted using SPSS 19.0 software. The matching tolerance was set at 0.1, and logistic regression was employed to fit independent and dependent variables, ensuring a balanced comparison between the two groups.
The benign lung tumors or tumor-like lesions included inflammatory pseudotumor, tuberculous ball, hamartoma, sclerosing hemangioma, fibroma, and inflammatory or infectious diseases. Malignant nodules were classified according to the 2015 World Health Organization pathological classification of lung cancer, including adenocarcinoma, squamous cell carcinoma, lymphoepithelioid carcinoma, and neuroendocrine tumor.
Thin-layer chest CT image data of the 360 nodules were imported into an AI analysis system (s-Discover/Lung, V1.0.2, 12-Sigma Technologies, USA). The system automatically performed nodule identification and benign-malignant analysis. The results were then reviewed by two senior physicians, who deleted false positives and verified the size, density, and location of the nodules. Two radiologists with over 10 years of experience in chest CT diagnosis independently identified pulmonary nodules and determined their benign or malignant characteristics. Statistical analysis was conducted using SPSS 19.0, with measurement data expressed as mean ± standard deviation. The receiver operating characteristic (ROC) curve was used to determine the acceptable malignant probability threshold, and the chi-square test was employed for group comparisons, with statistical significance set at P < 0.05.
The area under the ROC curve (AUC) for AI in diagnosing early-stage lung cancer was 0.771. At the optimal threshold of 85.5% malignant probability, the sensitivity and specificity of AI were 62.8% and 77.8%, respectively. While the sensitivity of AI (62.8%) was lower than that of radiologists (68.3%), its specificity (77.8%) was higher than that of radiologists (62.8%). Both differences were statistically significant (chi-square = 8.48, 6.96; P < 0.05). The diagnostic efficacy of AI and radiologists, evaluated using the Youden index and Kappa value, was not high (0.41 and 0.31, respectively).
The sensitivity of AI for nodules of different sizes and densities was consistently lower than that of radiologists, with no statistically significant differences observed (chi-square = 0.006, 3.174, 0.194; 0.029, 1.331, 3.669; P > 0.05). However, the specificity of AI for nodules of different sizes and densities was higher than that of radiologists, with statistically significant differences in the 5.0 to 10.0 mm and 10.1 to 20.0 mm groups (chi-square = 4.916, 5.733; P 0.05).
When AI and radiologists were combined, the sensitivity of the diagnosis significantly improved to 83.3% (chi-square = 60.72, 76.05; P < 0.05), while the specificity decreased to 52.8% (chi-square = 57.48, 119.28; P < 0.05). This suggests that the combination of AI and human expertise can enhance the detection of early-stage lung cancer, albeit with a trade-off in specificity.
The urgent clinical need for accurate and efficient diagnostic tools has driven the development of AI in pulmonary nodule detection and benign-malignant classification. Deep learning, a branch of AI, improves diagnostic accuracy by building learning models and conducting extensive training. Convolutional neural networks (CNNs), a common deep learning model, have shown better sensitivity and accuracy in classifying pulmonary nodules compared to radiologists in previous studies. However, this study highlights the limitations of AI in achieving high sensitivity, which may be attributed to the quality of the original database used for training.
In conclusion, AI based on deep learning demonstrates better specificity in diagnosing early-stage lung cancer but falls short in sensitivity compared to radiologists. The combination of AI and radiologists can achieve higher sensitivity, making it a valuable tool for the early diagnosis of lung cancer. Further research and large-scale clinical validation are needed to optimize AI models and improve their diagnostic performance.
doi.org/10.1097/CM9.0000000000000634
Was this helpful?
0 / 0