Predicting Invasiveness of Lung Adenocarcinoma from Chest CT with Few-shot Vision-Language Ternary Classification Model Authors: Nan Xu, Qianqian He, Lu Wang, Zhiwen Zhang, Qiuju Sheng, Shang Gao, Shimin Zhang, Bosinan Chen, Jianing Sun, Zhijian Zhang, Jie Zhang, Jing Qiu, Yunan Wang, Guanyu Liu, Enyu Li, Mingke Tian, Haotian Wang, Jiaping Yu, Yan Dong, Si Gao, Song Chen, Fan Yang, Zhihui Chang, Yue Dong, Lina Zhang, Jiangdian Song Published: 2025 ABSTRACT AbstractPreoperative differentiation of preinvasive lesions, minimally invasive adenocarcinomas, and invasive adenocarcinomas within pure ground-glass nodules (pGGNs) is challenging. Herein, this study investigated the potential of vision-language models to assist radiologists in noninvasively predicting pGGN invasiveness on CT scans. This retrospective multicenter study enrolled 848 patients with pathologically-confirmed lung adenocarcinoma manifesting as pGGNs. GPT-4o was tasked with localizing pGGNs on CT scans to detect ten pGGN invasiveness-associated features to generate a diagnosis and was compared with Molmo. The twenty-shot GPT-4o model demonstrated superior performance in the ternary classification of pGGN invasiveness (Delong test, P < 0.01). Six radiologists’ assessments revealed that GPT-4o output showed high reliability, willingness to use, reliance, low risk of harm, inappropriate content, and missing content. With GPT-4o assistance, another six radiologists achieved an average improvement in pGGN invasiveness diagnosis. The twenty-shot-based GPT-4o model exhibited superior diagnostic capability for pGGN invasiveness in lung adenocarcinoma, achieving significantly improved diagnostic accuracy by radiologists. Subject terms: Non-small-cell lung cancer, Diagnosis, Computational biology and bioinformatics, Communication and replication, Cancer imaging Introduction Ground-glass opacities are frequently observed as pulmonary abnormalities on chest computed tomography (CT) scans. Pure ground-glass nodules (pGGNs) are indicative of lung adenocarcinoma (LUAD). However, the underlying histology is diverse, ranging from inflammation and interstitial fibrosis to atypical adenomatous hyperplasia (AAH), adenocarcinoma in situ (AIS), primary lung cancer, and metastasis. Despite the indolent nature of pGGNs and current clinical guidelines recommending watchful waiting strategies1–3, invasive malignancy (1.7–24.3%) is detected on surgical resection4,5. Although prognosis is improved for patients with pGGN following treatments such as complete sublobar resection5, surgical management depends on the degree of pathological invasiveness, which ranges from pre-invasive lesions (AAH and AIS), minimally invasive adenocarcinoma (MIA), and invasive adenocarcinoma (IAC)6,7. Therefore, preoperative evaluation of LUAD invasiveness is crucial for informed clinical decision-making8. Given the limitations of biopsy and postoperative pathological evaluation in pGGN assessment, noninvasive imaging-based methods have been proposed as novel methods to predict pGGN invasiveness9–11. Radiomics and deep learning (DL) methods using preoperative CT have shown promise in predicting the pathological subtypes of LUAD12,13. Artificial intelligence (AI) techniques potentially differentiate preinvasive lesions, MIA, and IAC by decoding GGN features from CT images14–16. However, the inherent lack of solid components in pGGNs presents a challenge for current computer vision-based AI algorithms. Unlike experienced radiologists, these algorithms struggle to visually detect and interpret the subtle, yet crucial features associated with LUAD invasiveness, such as air bronchograms, vacuolization, vascular penetration, and pleural indentation1,17. CT features such as size and density, which are easily detected by the algorithm, often overlap among different GGN subcategories18. Furthermore, previous studies have mainly focused on the binary classification of invasiveness (e.g., AAH, AIS, and MIA vs. IAC, or AAH and AIS vs. MIA and IAC) within GGNs. Current ternary classification studies either do not specifically target pGGNs or face practical limitations in radiology because of the multi-model integration they require14,19–21. Consequently, clinicians have an urgent need for a method capable of directly identifying and describing key pGGN features from CT scans to enable accurate ternary classification of invasiveness in LUAD. Large language models (LLMs) have proven valuable in improving lung disease management22. Text- and image-based LLMs have demonstrated performances comparable to those of radiologists in generating standardized discharge summaries, radiology reports, and pathological classifications23–25. The latest iteration, GPT-4o, is a vision-language model capable of analyzing video data, offering a promising avenue for enabling AI to directly perceive key features associated with pGGN invasiveness from CT scans. GPT-4o can automatically localize lung nodules within CT scan videos, identify malignant characteristics, and achieve diagnostic accuracy of lung nodule malignancy with an accuracy comparable to that of traditional radiomics and DL methods26. These findings underscore the potential of GPT-4o in predicting LUAD invasiveness from pGGN CT images. This study proposes a novel few-shot vision-language ternary classification model to determine the optimal training strategy for GPT-4o in diagnosing pGGN invasiveness and evaluates its value in assisting radiologists in real-world diagnosis of pGGN. Results Patient characteristics Figure 1 illustrates the study workflow. In total, 848 patients (222 males, 26.18%, mean age [standard deviation, SD]: 56.32 [10.63] years) from four hospitals were included in this study (see patient enrollment flowchart in Fig. 2). The mean size (SD) of pGGNs was 10.68 (4.66) mm. A total of 333 patients had pathologically confirmed pre-invasive lesions, 376 had MIA, and 139 had IAC. Patient characteristics are presented in Table 1. A total of 848 CT videos (one video per pathologically confirmed pGGN per patient) were generated, with an average duration of 17.07 s per video. Overview of this study. Flowchart of patient inclusion and exclusion criteria for the four datasets. pGGN pure ground glass nodule, PRE preinvasive lesions, MIA minimally invasive lesions, IAC Invasive adenocarcinoma. Table 1Baseline characteristics of patients from the four centersVariableCenter1Center2Center3Center4Patients num274208216150Age (mean [SD])57.43 (10.21)56.84 (9.60)56.27 (10.77)53.61 (12.08)Age range23–7525–6321–7722–78Sex Male (%)81 (29.56)31 (14.90)61 (28.24)49 (32.67) Female (%)193 (70.44)177 (85.10)155 (71.76)101 (67.33)Pathology AAH/AIS (%)108 (39.42)88 (42.31)86 (39.81)51 (34.00) MIA (%)117 (42.70)87 (41.83)97 (44.91)75 (50.00) IAC (%)49 (17.88)33 (15.87)33 (15.28)24 (16.00)Nodule size (mean [SD])11.52 (5.24)10.12 (4.51)10.74 (4.59)9.81 (3.46)Location Right upper lobe (%)97 (35.40)81 (38.94)98 (45.37)51 (34.00) Right lower lobe (%)45 (16.42)36 (17.31)52 (24.07)26 (17.33) Right middle lobe (%)20 (7.30)9 (4.33)6 (2.78)8 (5.33) Left upper lobe (%)68 (24.82)53 (25.48)40 (18.52)46 (30.67) Left lower lobe (%)44 (16.06)29 (13.94)20 (9.26)19 (12.67)AAH atypical adenomatous hyperplasia, AIS adenocarcinoma in situ, MIA minimally invasive adenocarcinoma, IAC invasive adenocarcinoma. Baseline characteristics of patients from the four centers AAH atypical adenomatous hyperplasia, AIS adenocarcinoma in situ, MIA minimally invasive adenocarcinoma, IAC invasive adenocarcinoma. GPT-4o detection of CT features and providing diagnosis In terms of nodule boundary delineation, GPT-4o achieved a mean Dice coefficient of 0.76. The twenty-shot setting achieved the best diagnostic performance for LUAD invasiveness in pGGNs. GPT-4o achieved an averaged accuracy of 0.75 (95% confidence interval [CI]: 0.72–0.78), sensitivity of 0.74 (95% CI: 0.70–0.77), specificity of 0.86 (95% CI: 0.84–0.88), precision of 0.75 (95% CI: 0.71–0.78), F1 score of 0.74 (95% CI: 0.71–0.77), and area under the curve (AUC) of 0.80 (95% CI: 0.77–0.83) across the four centers, respectively (Table 2 and Supplementary Fig. 1). The result was significantly better than that of the other training strategies using <20 training samples (P < 0.01, Delong test, Supplementary Table 1). Although a slight improvement in prediction accuracy was observed compared with the thirty-shot training strategy, it was not statistically significant (P = 0.311, Delong test, Supplementary Table 1). GPT-4o’ s ternary classification performance for pGGN invasiveness was evaluated across different scanner manufacturers (Supplementary Table 2), slice thicknesses (Supplementary Table 3), tube currents (Supplementary Table 4), nodule sizes (Supplementary Table 5), and medical centers (Supplementary Table 6). Model performance was compared across three standard clinical window settings: (window width, window level) = (2000, −500), (2000, -600), and (1500, −500). The results indicated no significant differences in model performance across the different window width and window level settings (all P > 0.05; Supplementary Table 7). Furthermore, the AUC of Molmo was 0.45 (95% CI: 0.40–0.50), 0.51 (95% CI: 0.42–0.60), 0.47 (95% CI: 0.35–0.59), and 0.53 (95% CI: 0.45–0.61) on the four datasets, significantly lower than GPT-4o (P < 0.001). In addition, the diagnostic performance of GPT-4o, Gemini 2.5 Pro, and a previously published deep learning model was compared. GPT-4o outperformed Gemini 2.5 Pro (P = 0.001), whereas no statistically significant difference was observed between GPT-4o and the previously published deep learning model (P = 0.282). Table 2The diagnostic performance of GPT-4o on each center based on the twenty-shot training strategyC1C2C3C4Accuracy PRE0.81 (223/274)[0.76–0.86]0.84 (174/208)[0.78–0.88]0.83 (180/216)[0.78–0.88]0.84 (126/150)[0.77–0.89] MIA0.73 (201/274)[0.68–0.78]0.76 (159/208)[0.70–0.82]0.78 (169/216)[0.72–0.84]0.76 (114/150)[0.68–0.83] IAC0.90 (246/274)[0.86–0.93]0.90 (187/208)[0.85–0.94]0.93 (201/216)[0.89–0.96]0.91 (136/150)[0.85–0.95]Sensitivity PRE0.78 (84/108)[0.69–0.85]0.80 (70/88)[0.70–0.87]0.80 (69/86)[0.70–0.88]0.75 (38/51)[0.60–0.86] MIA0.69 (81/117)[0.60–0.77]0.72 (63/87)[0.62–0.81]0.77 (75/97)[0.68–0.85]0.77 (58/75)[0.66–0.86] IAC0.67 (33/49)[0.52–0.80]0.70 (23/33)[0.51–0.84]0.70 (23/33)[0.51–0.84]0.71 (17/24)[0.49–0.87]Specificity PRE0.84 (139/166)[0.77–0.89]0.87 (104/120)[0.79–0.92]0.85 (111/130)[0.79–0.91]0.89 (88/99)[0.81–0.94] MIA0.76 (120/157)[0.69–0.83]0.79 (96/121)[0.71–0.86]0.79 (94/119)[0.71–0.86]0.75 (56/75)[0.63–0.84] IAC0.95 (213/225)[0.91–0.97]0.94 (164/175)[0.89–0.97]0.97(178/183)[0.94–0.99]0.94 (119/126)[0.89–0.98]Precision PRE0.76 (84/111)[0.67–0.83]0.81 (70/86)[0.72–0.89]0.78 (69/88)[0.68–0.86]0.78 (38/49)[0.63–0.88] MIA0.69 (81/118)[0.59–0.77]0.72 (63/88)[0.61–0.81]0.75 (75/100)[0.65–0.83]0.75 (58/77)[0.64–0.84] IAC0.73 (33/45)[0.58–0.85]0.68 (23/34)[0.49–0.83]0.82 (23/28)[0.63–0.94]0.71 (17/24)[0.49–0.87]F1 score PRE0.77 (168/219)[0.70–0.83]0.80 (140/174)[0.74–0.87]0.79 (138/174)[0.72–0.86]0.76 (76/100)[0.66–0.85] MIA0.69 (162/235)[0.63–0.75]0.72 (126/175)[0.64–0.79]0.76 (150/197)[0.69–0.82]0.76 (116/152)[0.69–0.83] IAC0.70 (66/94)[0.58–0.80]0.69 (46/67)[0.54–0.80]0.75 (46/61)[0.62–0.86]0.71 (34/48)[0.56–0.85]See Supplementary Fig. 1 for the specific number in each predicted subtype; data in brackets are 95%Cls.C1 to C4 are the four centers, PRE preinvasive lesions, MIA minimally invasive lesions, IAC Invasive adenocarcinoma. The diagnostic performance of GPT-4o on each center based on the twenty-shot training strategy 0.81 (223/274) [0.76–0.86] 0.84 (174/208) [0.78–0.88] 0.83 (180/216) [0.78–0.88] 0.84 (126/150) [0.77–0.89] 0.73 (201/274) [0.68–0.78] 0.76 (159/208) [0.70–0.82] 0.78 (169/216) [0.72–0.84] 0.76 (114/150) [0.68–0.83] 0.90 (246/274) [0.86–0.93] 0.90 (187/208) [0.85–0.94] 0.93 (201/216) [0.89–0.96] 0.91 (136/150) [0.85–0.95] 0.78 (84/108) [0.69–0.85] 0.80 (70/88) [0.70–0.87] 0.80 (69/86) [0.70–0.88] 0.75 (38/51) [0.60–0.86] 0.69 (81/117) [0.60–0.77] 0.72 (63/87) [0.62–0.81] 0.77 (75/97) [0.68–0.85] 0.77 (58/75) [0.66–0.86] 0.67 (33/49) [0.52–0.80] 0.70 (23/33) [0.51–0.84] 0.70 (23/33) [0.51–0.84] 0.71 (17/24) [0.49–0.87] 0.84 (139/166) [0.77–0.89] 0.87 (104/120) [0.79–0.92] 0.85 (111/130) [0.79–0.91] 0.89 (88/99) [0.81–0.94] 0.76 (120/157) [0.69–0.83] 0.79 (96/121) [0.71–0.86] 0.79 (94/119) [0.71–0.86] 0.75 (56/75) [0.63–0.84] 0.95 (213/225) [0.91–0.97] 0.94 (164/175) [0.89–0.97] 0.97(178/183) [0.94–0.99] 0.94 (119/126) [0.89–0.98] 0.76 (84/111) [0.67–0.83] 0.81 (70/86) [0.72–0.89] 0.78 (69/88) [0.68–0.86] 0.78 (38/49) [0.63–0.88] 0.69 (81/118) [0.59–0.77] 0.72 (63/88) [0.61–0.81] 0.75 (75/100) [0.65–0.83] 0.75 (58/77) [0.64–0.84] 0.73 (33/45) [0.58–0.85] 0.68 (23/34) [0.49–0.83] 0.82 (23/28) [0.63–0.94] 0.71 (17/24) [0.49–0.87] 0.77 (168/219) [0.70–0.83] 0.80 (140/174) [0.74–0.87] 0.79 (138/174) [0.72–0.86] 0.76 (76/100) [0.66–0.85] 0.69 (162/235) [0.63–0.75] 0.72 (126/175) [0.64–0.79] 0.76 (150/197) [0.69–0.82] 0.76 (116/152) [0.69–0.83] 0.70 (66/94) [0.58–0.80] 0.69 (46/67) [0.54–0.80] 0.75 (46/61) [0.62–0.86] 0.71 (34/48) [0.56–0.85] See Supplementary Fig. 1 for the specific number in each predicted subtype; data in brackets are 95%Cls. C1 to C4 are the four centers, PRE preinvasive lesions, MIA minimally invasive lesions, IAC Invasive adenocarcinoma. Evaluation by radiologists Based on the evaluations of the six radiologists on feature output, GPT-4o achieved a mean (SD) of 4.43/5.00 [0.30] and a median (interquartile range [IQR]) of 4.40/5.00 [4.23-4.65] in reliability for assessing the ten features associated with pGGN invasiveness (Table 3). Specifically, six radiologists (R1–R6, with clinical experience of 5, 7, 11, 14, 17, and 20 years, respectively) rated the GPT-4o descriptions of feature “volume,” “margin,” and “shape” as closest to their own assessments across all datasets (mean [SD]: 4.75/5.00 [0.62], 4.65/5.00 [0.79], and 4.60/5.00 [0.85], and median [IQR]: 5.00/5.00 [5.00–5.00], 5.00/5.00 [5.00–5.00], and 5.00/5.00 [5.00–5.00], respectively). In contrast, GPT-4o descriptions relative to “vascular penetration” and “pleural indentation” features received the lowest concordance ratings compared with annotation by radiologists (mean [SD]: 4.01/5.00 [1.47] and 3.81/5.00 [1.65], and median [IQR]: 5.00/5.00 [3.00–5.00] and 5.00/5.00 [2.00–5.00]) (Table 3). Table 3Likert scale scores (mean with standard deviation [SD], median with interquartile range [IQR]) of the six radiologists for the ten featuresR1R2R3R4R5R6VolumeMean (SD)4.35 (0.70)4.77 (0.67)4.92 (0.39)4.66 (0.80)4.88 (0.50)4.94 (0.31)Median (IQR)4 (4–5)5 (5–5)5 (5–5)5 (5–5)5 (5–5)5 (5–5)MarginMean (SD)4.14 (0.85)4.58 (0.93)4.79 (0.85)4.82 (0.69)4.82 (0.61)4.76 (0.54)Median (IQR)4 (4–5)5 (5–5)5 (5–5)5 (5–5)5 (5–5)5 (5–5)ShapeMean (SD)4.07 (0.96)4.61 (0.85)4.68 (0.98)4.77 (0.71)4.75 (0.72)4.73 (0.62)Median (IQR)4 (4–5)5 (5–5)5 (5–5)5 (5–5)5 (5–5)5 (5–5)DensityMean (SD)4.02 (0.90)4.70 (0.71)4.79 (0.85)4.47 (1.08)4.80 (0.78)4.75 (0.66)Median (IQR)4 (3–5)5 (5–5)5 (5–5)5 (5–5)5 (5–5)5 (5–5)MassMean (SD)4.16 (0.82)4.67 (0.84)4.60 (1.14)4.45 (1.09)4.65 (1.07)4.88 (0.40)Median (IQR)4 (4–5)5 (5–5)5 (5–5)5 (5–5)5 (5–5)5 (5–5)Air bronchogram signMean (SD)4.19 (1.23)4.36 (1.18)4.89 (0.65)4.23 (1.49)4.87 (0.59)4.38 (1.33)Median (IQR)5 (4–5)5 (4–5)5 (5–5)5 (5–5)5 (5–5)5 (5–5)Internal vascular thickening or increaseMean (SD)3.99 (1.04)4.53 (0.91)4.77 (0.92)4.47 (1.23)4.81 (0.68)4.35 (1.21)Median (IQR)4 (3–5)5 (4–5)5 (5–5)5 (5–5)5 (5–5)5 (4–5)Internal vacuolizationMean (SD)4.02 (1.47)4.30 (1.16)4.56 (1.21)4.56 (1.25)4.54 (1.21)4.42 (1.34)Median (IQR)5 (3–5)5 (4–5)5 (5–5)5 (5–5)5 (5–5)5 (5–5)Vascular penetrationMean (SD)3.55 (1.49)3.67 (1.55)4.44 (1.36)3.92 (1.57)4.51 (1.11)3.98 (1.48)Median (IQR)4 (2–5)4 (2–5)5 (5–5)5 (3–5)5 (5–5)5 (2–5)Pleural indentationMean (SD)4.14 (1.21)3.74 (1.61)3.43 (1.94)4.22 (1.53)3.56 (1.80)3.79 (1.59)Median (IQR)5 (4–5)5 (2–5)5 (1–5)5 (5–5)5 (1–5)5 (2–5)R1 to R6 represent the six radiologists. Likert scale scores (mean with standard deviation [SD], median with interquartile range [IQR]) of the six radiologists for the ten features R1 to R6 represent the six radiologists. Regarding the clinical usefulness, evaluation by the six radiologists (R1-R6) demonstrated that they had high willingness to use GPT-4o output in the clinic (mean [SD]: 4.55/5.00 [0.77], median [IQR]: 5.00/5.00 [4.00–5.00]) and strong reliance of the feature output (mean [SD]: 4.22/5.00 [0.83]; median [IQR]: 4.00/5.00 [4.00–5.00]), and conceived a low risk of harm (mean [SD]: 1.36/5.00 [0.72], median [IQR]: 1.00/5.00 [1.00–2.00]), of inappropriate content (mean [SD]: 1.66/5.00 [0.87], median [IQR]: 1.00/5.00 [1.00–2.00]), or of missing content (mean [SD]: 1.42/5.00 [0.77], median [IQR]: 1.00/5.00 [1.00–2.00]). Overall, GPT-4o, which offers both diagnostic predictions and accompanying textual explanations, received the highest average ratings for willingness to use (4.55 vs. 2.13 vs. 3.46) and reliance (4.22 vs. 2.56 vs. 3.78) compared with Gemini 2.5 Pro and the deep learning model. In contrast, GPT-4o obtained the lowest ratings for perceived potential for harm (1.36 vs. 3.12 vs. 2.06) and missing content (1.42 vs. 2.36 vs. 2.66), demonstrating greater user confidence and perceived completeness. However, because GPT-4o generates detailed narrative descriptions and visual evidence rather than a single malignancy probability, radiologists noted a slightly higher risk of inappropriate or excessively detailed content compared with Gemini 2.5 and the deep learning model (1.66 vs. 1.35 vs. 1.12). Reader study and reproducibility evaluation Without GPT-4o assistance, another six radiologists (R7–R12, with 3, 3, 5, 5, 7, and 12 years of experience, respectively) achieved accuracies of 0.34, 0.45, 0.65, 0.40, 0.59, and 0.67, respectively, in diagnosing 110 patients, which was inferior to the diagnostic accuracy of GPT-4o (0.73). After a four-week washout period, the diagnostic accuracies of the six radiologists with GPT-4o assistance for the same patients improved to 0.62 (28% increase), 0.65 (20%), 0.76 (11%), 0.69 (29%), 0.83 (24%), and 0.86 (19%), respectively, compared with the first-round diagnosis (McNemar’s test, P < 0.001, Fig. 3). In the evaluation of 110 nodules by six radiologists, GPT-4o assistance resulted in 98 diagnostic corrections in the high-Dice group and 71 in the low-Dice group, representing a statistically significant difference (P = 0.02). Among the cases correctly classified by GPT-4o but misdiagnosed by senior radiologists, seven patients were identified (three with MIA and four with IAC). Specifically, the three MIA nodules were categorized as preinvasive lesions by both radiologists, while three IAC nodules were downgraded to MIA, and one IAC nodule was interpreted as a preinvasive lesion. Conversely, ten patients (five with preinvasive lesions, four with MIA, and one with IAC) were correctly diagnosed by radiologists but misclassified by GPT-4o. In this subgroup, all five preinvasive lesions were overestimated as MIA by GPT-4o, four MIA lesions were predicted as IAC, and one IAC lesion was misclassified as MIA. Comparison of the six radiologists’ diagnosis accuracy with and without the GPT-4o’s assistance. R7 to R12 represent the six radiologists with 3, 3, 5, 5, 7, and 12 years of experience. For the reproducibility evaluation based on two rounds of testing, GPT-4o demonstrated a reproducibility of 92.00% (46/50) for the three-class diagnosis of pGGN invasiveness. The assessment of GPT-4o feature descriptions by three radiologists yielded an average reproducibility of 81.87%. Specifically, GPT-4o feature reports of “volume,” “margin,” and “shape” showed the highest reproducibility (87.98%, 87.23%, 84.66%), whereas feature “vascular penetration” and “pleural indentation” had the lowest reproducibility (72.34% and 77.39%). The best and worst features in terms of the feature reliability of the GPT-4o output, as assessed by radiologists, were consistent with the assessment of feature reproducibility by radiologists. Discussion This study showed that GPT-4o can automatically detect and report key nodule features associated with pGGN invasiveness in CT images. This study demonstrated the optimal GPT-4o training strategy and further showed that the accuracy of GPT-4o in diagnosing the invasiveness of pGGN in LUAD exceeded that of other vision-language models, such as Molmo. In addition, the assessment of GPT-4o output across ten pGGNs invasiveness-associated features by six radiologists across multiple datasets revealed strong agreement with the radiologists’ manual assessment, suggesting its value in assisting pGGNs feature detection and report in radiology. Moreover, GPT-4o significantly improved the ability of radiologists to diagnose pGGN invasiveness in real-world practice, yielding 24%, 20%, 24%, and 19% improvement for radiologists with 3, 5, 7, and 12 years of experience, respectively (McNemar’s test, P < 0.001). The variation in performance improvement with GPT-4o assistance among radiologists of different experience levels appears to reflect their differing patterns of model utilization. Junior radiologists tended to rely more on GPT-4o’s feature reports, which they found particularly valuable for recognizing subtle imaging signs such as the air bronchogram. In contrast, senior radiologists primarily used GPT-4o as a supplementary tool to corroborate their own assessments, especially in patients with ambiguous or complex imaging findings. Under the assistance of GPT-4o, radiologists achieved more diagnostic corrections in the high-Dice group than in the low-Dice group (P = 0.02), suggesting that radiologists benefit more from GPT-4o assistance when the automatically generated contours are more accurate. Through the review of discordant cases between GPT-4o and radiologist diagnoses, it was observed that GPT-4o tends to overestimate the invasiveness of lung nodules, whereas radiologists generally adopt a more conservative diagnosis. When invasive features are subtle, GPT-4o often overinterprets them as signs of invasiveness, while radiologists remain more cautious in their assessments. CT features that are closely related to LUAD invasiveness, such as air bronchograms, vacuolization, and vascular penetration, require qualitative assessment by radiologists via visual inspection. However, these features are difficult to describe using current DL networks, which typically provide only a single probability value, thereby limiting their clinical reliability27. A three-class prediction of tumor invasiveness in sub-centimeter pulmonary adenocarcinomas has been reported using neural networks14; however, the above unique characteristics of pGGNs were not considered. Conversely, due to the lack of solid components and the inherent early-stage nature of pGGNs, determining histological subtypes based on traditional CT features such as size and density is unreliable9. The study demonstrates that GPT-4o potentially overcomes the limitations of traditional DL methods and mimics the visual assessment of CT features in pGGNs by radiologists. In the evaluation by six radiologists, the GPT-4o reported features of “volume,” “margin,” and “shape,” which were identified as highly-concerned morphological features of pGGNs in the assessment of LUAD invasiveness, achieved the highest agreement with the diagnoses made by the radiologists4,5,18,28,29. Overall, the agreements of the features “vascular penetration” and “pleural indentation” between GPT-4o and radiologists were lower than the other features, indicating that GPT-4o’s ability to identify tiny features in pGGN requires further improvements. Regarding vascular penetration, senior radiologists reported that traversal by small vessels was difficult for GPT-4o to distinguish, whereas large vessels were sometimes misclassified as solid components. For pleural indentation, they observed that subtle pleural traction—particularly involving the interlobar pleural surface and minor deformations such as perinodular fibrotic-strand retraction or mild indentation in pleura-abutting nodules—remained challenging for GPT-4o to accurately identify. Specifically, it is necessary to detect the difference between the blood vessels in the nodule and the similar pGGN components in CT images before determining the vascular penetration. In addition, the cause of pleural indentation may be the proliferation of fibrous tissue in the internal pGGNs, as well as the depression of the visceral pleura and thickening of the pleural adhesions caused by atelectasis. However, these imaging features are more difficult to detect in pleural adhesions than the traditional morphological features17,30,31. Preinvasive lesions, MIA, and IAC have different clinical outcomes32. For invasive adenocarcinoma, lobectomy or segmentectomy should be considered to prevent malignant progression6,33, whereas wide-wedge resection without lymph node dissection is sufficient for AIS and yields a better prognosis. Although watchful waiting remains the recommended approach for preinvasive lesions1–3, the importance of determining an appropriate treatment strategy at the time of pGGN detection has been highlighted5. Given the limitations of biopsy or postoperative pathology in pGGN assessment, noninvasive imaging-based methods have been proposed for predicting pGGN invasiveness9–11. This study is superior to previous studies in demonstrating the high reliability, completeness, and harmlessness of GPT-4o output as assessed by six radiologists, highlighting its potential value, particularly in providing textual descriptions of detected CT features. This study is an important advance as it identifies the optimal training strategy for predicting the three-class invasiveness of pGGNs using GPT-4o in both zero- and few-shot settings. A key distinction between fine-tuning LLMs and traditional deep learning methods lies in the requirement of limited training data for LLMs. Contrary to conventional wisdom that model accuracy improves with increasing training data, these findings reveal a nuanced relationship within this few-shot learning paradigm. The model accuracy increased (from 0.57 to 0.79, P < 0.001, Supplementary Table 1) as the number of training samples increased from zero-shot to twenty-shot. However, further increases in training data (thirty-shot) did not yield additional performance gains (0.73 to 0.75, P > 0.05), suggesting that twenty-shot learning represents the optimal approach for GPT-4o in predicting pGGNs invasiveness. Although previous studies have demonstrated the efficacy of two-shot learning with the GPT-4V for medical image analysis34, the dynamic nature of video analysis indicated that future research on video-based tasks using GPT should consider training sample sizes closer to twenty-shot to ensure an accurate analysis of key features. Conventional deep learning models generally achieve high diagnostic accuracy when trained on large, well-curated datasets. However, they often lack clinical interpretability and require professional knowledge in computer science (such as building neural networks), which limits their flexibility in real-world clinical settings. In contrast, GPT-4o can achieve comparable diagnostic performance with only a few training samples through few-shot learning. This process is more user-friendly for clinicians who lack knowledge in computer science than the deep learning models. It also provides explicit reasoning and feature-based explanations, thereby improving decision transparency and interpretability for radiologists. In addition, we also noticed that although GPT-4o generates more extensive textual and visual outputs, this information may introduce visual or cognitive distractions, potentially leading to misinterpretation. These complementary characteristics suggest that traditional DL models and vision–language models like GPT-4o could be integrated in future workflows by combining the quantitative precision of DL with the explanatory transparency of GPT-4o to enhance clinical trust and usability. This study had some limitations. First, this study provided GPT-4o with the coordinates of pGGNs in CT scans based on previously reported limitations in automated lung nodule detection26. Although this pipeline reduces the requirement for manual intervention compared with previous DL methods (such as tumor region clipping)8,14,20, future research should improve the performance of LLMs for the automatic detection of pGGN on CT images. Second, while this study included four datasets with 848 patients, further validation using a wider range of datasets, including a more diverse group of patients with varying baseline characteristics, is necessary to confirm the generalizability of GPT-4o. All patient-identifying information was deleted upon converting CT images into videos; however, potential privacy leakage risks remain when using online models for analysis. Therefore, the transparency of the GPT-4o algorithm must be highlighted before clinical translation. To further protect patient privacy, both researchers and developers share responsibility. Researchers should ensure rigorous dataset de-identification and make anonymization protocols publicly accessible, while developers should adopt privacy-preserving designs such as encryption, local data processing, and transparent data usage policies. Mutual transparency and accountability can help large-model applications advance within an ethical, secure, and reproducible framework for medical AI. Finally, owing to the limited availability of LLMs capable of video analysis at the time of writing, this study included only GPT-4o and Molmo35. Future studies should compare the performance of other vision-language models in predicting LUAD invasiveness. Future work will aim to incorporate benign GGNs, such as inflammatory or fibrotic lesions, to capture the full radiological spectrum of pGGOs and enhance the model’s clinical applicability. In addition, cine CT may provide valuable insights by capturing subtle respiratory motion and dynamic changes in nodule morphology; therefore, future studies should further explore its role in comprehensive pGGO evaluation. Finally, incorporating more extensive longitudinal follow-up data will allow for a more robust assessment of vision LLM in analyzing temporal changes in GGNs. In conclusion, this study presents a reliable tool able to assist radiologists in accurately diagnosing the pGGN invasiveness in LUAD preoperatively using readily available CT images. This study determined the optimal training strategy for GPT-4o to perform this task and demonstrated its value in improving radiological practice. With further external validation and transparent regulation, these findings inform the potential of GPT-4o to contribute to better clinical decision-making regarding pGGN invasiveness. Methods Study design This study first input prior knowledge of CT features significantly correlated with three invasive subtypes of pGGN on CT images into GPT-4o for memorization. The CT videos of each patient, composed of sequential images, were individually input into GPT-4o to perform pGGN invasiveness classification. The accuracy of the three-class predictions of GPT-4o was evaluated by comparing its diagnosis with pathological results. In addition, GPT-4o textual descriptions of ten pGGN-related CT features were compared with annotations by radiologists across different datasets to assess the reliability of the GPT-4o output. Finally, a reader study was conducted to determine the improvement in diagnostic accuracy by radiologists in classifying pGGN invasiveness with the assistance of GPT-4o. Patients This retrospective study included patients diagnosed with LUAD at four hospitals between 1 September 2014 and 1 September 2024. Inclusion criteria were: (1) age ≥18 years with complete demographic data, (2) availability of preoperative chest CT scans (Details of the CT scanners used across different centers are provided in Supplementary Note 1), (3) pGGN determined by two radiologists (≥10 years of experience) according to established guidelines without a solid component, and (4) All preinvasive lesions and invasive subtypes were pathologically confirmed through biopsy or surgical resection to ensure diagnostic consistency and labeling accuracy across the dataset. The exclusion criteria were as follows: (1) prior diagnosis of other cancers or surgery, (2) non-primary LUAD, and (3) poor CT image quality or missing baseline clinical data. This multicenter retrospective study adhered to the principles of the Declaration of Helsinki and approved by the Institutional Review Board and Ethics Committee of Shengjing Hospital of China Medical University (2024PS1395K), The Fourth Affiliated Hospital of China Medical University (2024-KS-151), Liaoning Cancer Hospital (XJS-20240460) and Zhuhai People's Hospital (2025-KT-122). All data were anonymized to protect patient privacy, and informed consent was waived, as the study involved retrospective analysis of existing data without any direct patient interaction. Chest CT processing Chest CT images of patients were sequenced in abdominal-cervical order. Based on previous studies, a frame rate of 20 frames per second was used in the present study26. The code used to generate the anonymized patient videos is provided in Supplementary Note 2 and all personal identification information was removed from the videos. CT image window width and level were set at 2000 and -500, respectively, to highlight specific nodule details. The resulting CT video had a resolution of 512×512×N pixels, where N represents the number of CT slices. All anonymized video data (848 patients) have been publicly released for research use (Mendeley Data: https://data.mendeley.com/datasets/h7b4ryrbzw/1). Prompt engineering To enable GPT-4o to learn the CT features of the three different pGGN invasiveness subtypes, this study first identified ten features associated with pGGN invasiveness based on current clinical guidelines and previous studies1,4,17. These features included nodule: (1) volume, (2) mass, (3) density, (4) shape, (5) margin, (6) internal vacuolization, (7) internal air bronchogram sign, (8) vascular penetration, (9) internal vessel thickening or increase, and (10) pleural indentation. These features formed textual prior knowledge for GPT-4o to memorize the pGGN features, as detailed in Supplementary Note 3. To determine the optimal training strategy for GPT-4o, this study employed zero-shot, one-shot, three-shot, five-shot, ten-shot, twenty-shot, and thirty-shot settings. Training samples were randomly selected from the study datasets (the number of preinvasive lesions, MIA, and IAC samples in the training dataset were set to maintain a ratio of approximately 40%: 40%: 20%, as reported in previously)14, and their corresponding CT videos and prompts (Supplementary Note 4) were input into GPT-4o to facilitate the memorization of the image characteristics for each pGGN invasiveness subtype. GPT-4o detection of CT features and providing a diagnosis After inputting the textual prior knowledge of pGGN features and training samples into GPT-4o, a step-by-step approach was adopted to guide the GPT-4o analysis for each patient (The video used for training GPT-4o was provided in Supplementary Video 1). First, the lesion in the video was automatically located based on the pGGN coordinates provided in the prompt (Supplementary Note 5). GPT-4o was then used to segment the pGGN and to draw the nodule boundary on the CT images based on the prompt (The specific prompts for detecting nodule boundaries are provided in Supplementary Note 5, and the corresponding video to illustrate the detection procedure is provided in Supplementary Video 2). The pGGN boundary segmentation generated by GPT-4o was compared with the manual segmentations performed by two radiologists, and the Dice coefficient was calculated to evaluate segmentation accuracy (The detailed procedure for calculating the Dice coefficient is provided in Supplementary Note 6). Two radiologists visually assessed the pGGN segmentation by GPT-4o to ensure that all pGGNs were correctly recognized. Finally, the GPT-4o was instructed to generate reports for each of the ten pGGN features detected within the identified nodule area, followed by a three-class diagnosis based on the detected features (preinvasive lesions, MIA, or IAC) accompanied by the predicted probability for each category. Evaluation by radiologists Six radiologists (R1–R6) were recruited to evaluate the reliability of the GPT-4o reports (Details of the evaluation are provided in Supplementary Note 7). Each radiologist independently reviewed the CT scans of each patient to annotate the ten CT features and was blinded to the GPT-4o output. Subsequently, they were asked to rate the concordance between their feature descriptions and those of GPT-4o using a 5-point Likert scale for each feature (1: completely incorrect; 2: more incorrect than correct; 3: equally correct and incorrect; 4: more correct than incorrect; and 5: completely correct) to measure the reliability of GPT-4o feature detection. Reader study An ad hoc reader study was conducted to evaluate the potential of GPT-4o to assist radiologists in the diagnosis of pGGN invasiveness. This study involved six additional radiologists (R7–R12) and 110 randomly selected patients (50 with pre-invasive lesions, 40 with MIA, and 20 with IAC). These radiologists initially diagnosed patients based on CT images and baseline information without GPT-4o assistance. After a four-week washout period, they reviewed the data again, accessing the feature report and diagnosis results of GPT-4o. By comparing the diagnostic accuracy of the two rounds, this study determined whether GPT-4o assistance could improve the accuracy of pGGN invasiveness diagnoses by radiologists (Details of the diagnostic improvements of radiologists with GPT-4o assistance are provided in Supplementary Note 7). Comparison of GPT-4o with other vision-language models A state-of-the-art open-source vision-language model, Molmo-7B, was used as a base point to further compare the performance of GPT-4o in pGGN invasiveness prediction on the four evaluated datasets. Molmo models are trained on PixMo, a dataset comprising one million highly curated image-text pairs, and have demonstrated near state-of-the-art performance on 11 academic benchmarks, ranking second only to GPT-4o35. This utilized the latest version, “Molmo-7B-D-0924,” in the experiments. The same inputs, including patient CT image videos and the corresponding prompts, were input into the Molmo-7B model to output the three-class diagnosis result. Additionally, GPT-4o was compared with Gemini 2.5 Pro and a previously published deep learning model21 trained on 4929 nodules from 4483 patients across multiple centers. The deep learning model was trained and inferred using an NVIDIA GeForce RTX 3090 GPU running Ubuntu 18.04 LTS. To further assess the clinical applicability of GPT-4o, Gemini 2.5 Pro, and the deep learning model, six radiologists (R1–R6) evaluated the willingness to use, reliance, perceived potential for harm, inappropriate content, and missing content of the outputs from the three models for each patient, using a five-point scale (1 = low; 5 = high). Statistical analysis To evaluate the performance of GPT-4o in pGGN invasiveness, the accuracy, sensitivity, specificity, precision, and F1 score were calculated. In addition, to compare the accuracy of GPT-4o with different training strategies, and between GPT-4o and radiologist diagnoses (in the reader study), the DeLong test was used to compare the AUC for each method36. Agreement for the feature detection between GPT-4o and the six radiologists (R1-R6) was assessed using Likert scale scores, reported as median and IQR, and mean and SD. McNemar’s test was used to evaluate the effects of GPT-4o assistance on diagnosis by clinicians by comparing diagnostic accuracy with and without GPT-4o assistance. In addition, a test-retest experiment was conducted to evaluate the reproducibility of the GPT-4o. Fifty patients (20 with preinvasive lesions, 20 with MIA, and 10 with IAC) were randomly selected for two rounds of repeated experiments (2-week intervals). The GPT-4o reports and diagnoses for each patient in both rounds were recorded. For the three-class diagnoses, GPT-4o reproducibility was determined by directly comparing the outputs from the two rounds. For textual feature reports, three additional radiologists with 5 years of experience independently assessed the consistency of GPT-4o’s feature descriptions between the two rounds using a predefined scale (range, 0–100%, with higher values indicating better consistency). All the experiments were conducted using the chat interface of the official OpenAI website, which allows video uploads. All statistical analyses were performed using R software (v.4.0.3; R Foundation for Statistical Computing). Statistical significance was defined as a two-sided P < 0.05. Supplementary information Supplementary materials Supplementary Video1 Supplementary Video2 Acknowledgements This work was funded by the National Natural Science Foundation of China (92259104), Scientific Research Fund of Liaoning Provincial Education Department (LJ242410159058), and the Xingliao Talent Program of Liaoning Province (No. XLYC2403102). The funders played no role in study design, data collection, analysis and interpretation of data, or the writing of this manuscript. Author contributions Study concepts and design: N.X., Q.H., L.W., Z.Z., Q.S., S.G., S.Z., Y.D., L.Z., J.S. Literature research: N.X., Q.H., S.G., S.Z., B.C., J.S., Z.Z., F.Y., J.S. Clinical studies: N.X., Q.H., L.W., Z.Z., Q.S., S.G., S.Z., B.C., J.S., Z.Z., J.Z., J.Q., Y.W., G.L., E.L., M.T., H.W., J.Y., Y.D., S.G., S.C. Experimental studies/data analysis: N.X., Q.H., L.W., Z.Z., Q.S., S.G., S.Z., B.C., J.S., Z.Z., J.Z., S.C., F.Y. Statistical analysis: N.X., Q.H., S.G., S.Z., S.C., F.Y., Z.C., J.S. Manuscript preparation: N.X., Q.H., L.W., Z.Z., Q.S., S.G., S.Z., F.Y., Z.C., L.Z., J.S. Manuscript editing: N.X., Q.H., L.W., S.C., F.Y., Z.C., Y.D., L.Z., J.S. All authors have read and approved the manuscript. Data availability The CT videos have been uploaded to Mendeley Data: https://data.mendeley.com/datasets/h7b4ryrbzw/1. The data that support the findings of this study are available from the corresponding authors with a signed data access agreement. Code availability The code for converting raw CT images into anonymized CT videos is available at Supplementary Note 2. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Nan Xu, Qianqian He, Lu Wang, Zhiwen Zhang, Qiuju Sheng, Shang Gao, Shimin Zhang. Contributor Information Song Chen, Email: chensongchina@163.com. Fan Yang, Email: fanyang@sdu.edu.cn. Zhihui Chang, Email: changzh@sj-hospital.org. Yue Dong, Email: dyy1026@sina.com. Lina Zhang, Email: lnzhang@cmu.edu.cn. Jiangdian Song, Email: song.jd0910@gmail.com. Supplementary information The online version contains supplementary material available at 10.1038/s41746-025-02229-2. References Associated Data Supplementary Materials Supplementary materials Supplementary Video1 Supplementary Video2 Data Availability Statement The CT videos have been uploaded to Mendeley Data: https://data.mendeley.com/datasets/h7b4ryrbzw/1. The data that support the findings of this study are available from the corresponding authors with a signed data access agreement. The code for converting raw CT images into anonymized CT videos is available at Supplementary Note 2.