Page 104 - TD-3-3

Tumor Discovery Effectiveness of AI imaging for lung nodules

qualitative diagnosis of lung nodules and the pathological carcinoma, neuroendocrine carcinoma, and small cell
results, which served as the gold standard. To assess the carcinoma. Inflammatory pseudotumors predominantly
diagnostic performance of both the AI and physician characterized benign nodules, while lung adenocarcinoma
readings in identifying benign and malignant lung was the most common malignant nodule. We present the
nodules, we employed the receiver operating characteristic detailed statistics of pathological diagnoses in Table 2.
(ROC) curve. Sensitivity and specificity were calculated We compared AI-assisted diagnostic results with
based on the optimal cutoff value determined from the the pathological examination results. Among the 226
ROC analysis.
malignant lung nodules confirmed by pathology, AI
3. Results correctly diagnosed 204 as malignant, misclassified 21 as
benign, and failed to recognize one nodule (located near
3.1. Basic clinical information the hilum). Among the 65 benign lung nodules confirmed
We surgically removed and pathologically examined a total by pathology, AI accurately diagnosed 35 as benign,
of 291 lung nodules. Of these, 65 were benign, and 226 misclassified 27 as malignant, and failed to recognize three
were malignant. The nodules included 118 solid nodules, nodules (located near the hilum). The Kappa value for the
102 partially solid nodules, and 71 pure ground-glass agreement between AI diagnoses and pathological results
(nonsolid) nodules. The diameters of the nodules varied as was 0.4883 (P < 0.001), indicating moderate agreement.
follows: 0 – 10 mm in 59 cases, 11 – 20 mm in 149 cases, and This result suggests that the qualitative diagnosis of lung
21 – 30 mm in 83 cases. We found no statistically significant nodules by AI generally aligns with pathological results
differences between benign and malignant nodules (Table 3).
regarding gender and nodule size (P > 0.05). However, the We analyzed the consistency between the diagnostic
number of malignant nodules was significantly higher than results from the physician-reading group and the
that of benign nodules among both partially solid nodules pathological examination results. In the group of
(P < 0.05) and nonsolid nodules (P < 0.05). Furthermore, 226 malignant lung nodules confirmed by pathology,
malignant nodules were significantly more prevalent in the physicians diagnosed 188 as malignant, misclassified 11
lobes superior (P < 0.05) (Table 1).
as benign, and left 27 nodules unclassified. In the group
3.2. Comparison of the efficacy of the AI diagnostic of 65 benign lung nodules confirmed by pathology,
system and roentgenologist in identifying benign physicians diagnosed 29 as benign, one as a calcified
and malignant lung nodules nodule, misclassified 22 as malignant, and left 13 nodules
unclassified. The Kappa value for the agreement between
This study included a total of 291 lung nodules that met the physician diagnoses and pathological results was 0.5581
criteria. All patients underwent thoracoscopic resection, (P < 0.001), indicating moderate agreement. However, a
which included wedge resection, segmentectomy, or significant number of nodules were left unclassified by the
lobectomy. We examined post-operative tissues using physicians (Table 4).
pathological methods. The results of the pathological
examination served as the gold standard for this study. We plotted ROC curves based on the malignant
According to pathological diagnosis, benign nodules probability of lung nodules predicted by the AI-assisted
comprised inflammatory pseudotumor, hamartoma, diagnostic system and the accuracy of physician readings.
organized pneumonia, tuberculosis, bronchogenic cyst, The area under the ROC curve (AUC) for the AI group
sclerosing pneumocytoma, and mycotic infection. Malignant was 0.727, with a sensitivity of 90.27% (95% confidence
nodules included adenocarcinoma, squamous carcinoma, interval [CI]: 0.8563 – 0.9380) and a specificity of 58.46%
poorly differentiated carcinoma, lymphoepithelioma-like (95% CI: 0.4556 – 0.7056). For the physician-reading

Table 1. Basic demographic and clinical information of subjects
Categorization Sex Average Nodule size Nature of the nodule Location of the nodule
age (diameter, mm)
Male Female 0 – 10 11 – 20 21 – 30 Solid Part‑solid Pure Lobus Lobus Lobus inferior
ground‑ superior middle
glass
Benign lesions 28 37 57.61 17 34 14 47 10 8 27 9 29
Malignant lesions 98 128 59.86 42 115 69 71 92 63 145 24 57
P-value 1 0.2372 2.428e-08 0.003591

Volume 3 Issue 3 (2024) 4 doi: 10.36922/td.4178