Background Although traditional cytopathological diagnosis of lung cancer has its advantages, it is greatly influenced by doctors' subjective experience and workload. The new generation of artificial intelligence, represented by deep learning algorithm models, can automatically extract and summarize features from medical images, demonstrating significant advantages in intelligent diagnosis.
Objective To develop an artificial intelligence cytopathological diagnosis system for lung cancer and explore its diagnostic value by combining the artificial intelligence and digital pathology.
Methods From May 2021 to July 2023, 533 patients with suspected lung cancer were selected from the First Medical Center of Chinese PLA General Hospital. Among them, 354 cases were finally diagnosed with lung cancer (including 98 cases of adenocarcinoma, 140 cases of squamous carcinoma, and 116 cases of small-cell carcinoma), and another 179 cases were non-lung cancer. The bronchoscopic biopsy specimens and pleural effusion specimens from the selected cases were smeared, stained, and scanned. Using the digital pathological slices from 340 randomly selected samples (including 229 lung cancer cases and 111 non-lung cancer cases), the candidate detection models and classification models were trained, validated, and tested, respectively. Based on the test results, the YOLO v7 detection model and Vision Transformer classification model were selected as the basic structure to initially establish the Artificial Intelligence Cytopathological Diagnosis System for lung cancer. Then the trained Artificial Intelligence Cytopathological Diagnosis System was used to diagnose the remaining 193 untrained samples for validation, and the interpretation results were compared with the pathological diagnosis results as the standard.
Results The accuracy of the developed Artificial Intelligence Cytopathological Diagnosis System in lung cancer diagnosis was 91.2% (176/193), with sensitivity of 98.4% (123/125), specificity of 77.9% (53/68), positive predictive value of 89.1% (123/138), and negative predictive value of 96.4% (53/55). The Youden index was 0.763, and Kappa statistic was 0.798.
Conclusion The Artificial Intelligence Cytopathological Diagnosis System has high sensitivity and accuracy in the diagnosis of lung cancer, which can effectively improve the efficiency of lung cancer diagnosis. However, the system still needs to be further optimized to enhance its diagnostic specificity.
DownLoad: