Page 103 - TD-3-3

Tumor Discovery Effectiveness of AI imaging for lung nodules

between 27 and 83 years, with a mean age of 57.07 years. medium-risk nodules (yellow) when the probability was
The study encompassed 291 lung nodules, categorized between 50% and 70%, and low-risk nodules (green)
into 65 benign nodules and 226 malignant nodules based when the probability was below 50%. In this research, we
on pathological findings. Among the malignant nodules, categorized lung nodules as benign (malignant probability
215 were diagnosed as lung adenocarcinoma. This study <70%) or malignant (malignant probability ≥70%) based
was reviewed and approved by the ethics committee of on the malignant probability. Any AI diagnostic results
our hospital (Ethics Approval Number 2023 Research No. that were questionable were reviewed by professional
[66]), and all enrolled patients provided informed consent. imaging physicians on the project team.
2.2. Pathologic diagnostic methods User Workflow (InferViewer v1.0.6 System): Doctors
upload imaging data through the system interface, which
The results of the pathological examination were considered then automatically performs lesion identification and
the gold standard in this study. All subjects underwent analysis. Users can view the automatically generated lesion
either trans-thoracic bronchoscopic puncture biopsy or list, which includes various attributes and classifications of
lung nodule resection. We directed the obtained specimens the nodules. The system provides a function to compare
to the pathology department for examination to determine images from different time points, aiding doctors in
the nature of the lung nodules and the pathological type of assessing changes in the condition. Doctors can select
lung cancer. We recorded the results of these pathological different report templates to generate structured reports
examinations. or use additional application tools for further analysis.

2.3. LDCT scanning methods They can also provide feedback on the system’s predictions
by marking missed nodules or removing false-positive
We followed the Chinese Lung Cancer Low-Dose Spiral CT lesions, which helps to enhance the system’s accuracy and
Screening Guidelines (2018 edition) for image acquisition reliability.
using a Philips 320-row and 640-slice CT scanner. The
parameters for the LDCT scans included a tube voltage of AI Software Processing Workflow (InferViewer
100 – 120 kV, a tube current of <70 mAs based on body v1.0.6 System): The software employs AI, specifically
mass index, a field of view of 500 mm, and a minimum layer deep learning, to automatically identify and analyze lung
thickness of 0.5 mm. We reconstructed the raw data into a nodules in CT images. This includes automatic annotation,
thin layer with a thickness ranging from 0.5 – 1.25 mm. segmentation measurement, nature classification, and
location pinpointing. The system features an intelligent
2.4. Methods of reading images follow-up option that monitors disease progression and
provides re-examination suggestions by comparing a
We used both physician reading and an AI method to patient’s medical imaging information over different
diagnose the CT images. The physician-reading group periods. Based on this information, the system can
included two attending physicians with 10 years of generate structured image-text reports. Users can view
experience in diagnosing chest images. They independently the automatically generated lesion list, which includes
analyzed and diagnosed the lung nodules according to various attributes (such as size and density, lobulation
the “Chinese Expert Consensus on the Diagnosis and sign, spiculation sign, vascular convergence sign, air
Treatment of Lung Nodules (2018 edition).” They classified bronchogram sign, and pleural retraction sign) and
lung nodules into six categories based on the Lung Imaging classifications of nodules. The image workstation offers
Reporting and Data System (Lung-RADS), as formulated a variety of application tools, including length and angle
by the American College of Radiology. In this study, we measurements, to support comprehensive observation and
considered lung nodules categorized as two and three to multi-angle assessment of lesions by doctors.
be benign, while those categorized as four and above were
considered malignant. We resolved any disagreements 2.5. Statistical analysis
through discussion or consultation with a third investigator.
We used SPSS 24.0 software exclusively for statistical
The AI reading group used the InferViewer v1.0.6 analysis. Measurement data were expressed as mean ±
system for assisted diagnosis. We input images (DICOM standard deviation (x- ± s) if normally distributed, or
format) of chest CT scans into the AI system, which then as median (interquartile range) if not. Count data were
identified and calculated the average diameter, density reported as frequency and percentage, with qualitative
classification, volume, and malignant probability of the data also expressed as percentages. We tested frequencies
lung nodules. The malignant probability ranged from and percentages using the chi-square test, considering
0 to 100%. The AI system indicated high-risk nodules P < 0.05 as statistically significant. We used the Kappa
(red) when the malignancy probability exceeded 70%, consistency test to evaluate the agreement between the AI’s

Volume 3 Issue 3 (2024) 3 doi: 10.36922/td.4178