Abstract 132P
Background
Radiomics, a quantitative imaging analysis technique powered by artificial intelligence, holds promise for lung cancer diagnosis. However, conventional radiomic diagnosis models are often based only on image-based features, potentially limiting their accuracy. This study assesses the impact of integrating clinical information into a deep-learning model for determining the malignancy of pulmonary nodules.
Methods
A cross-sectional study of 87 resected nodules from 90 patients was conducted. Clinical data included demographic variables, epidemiological risk factors, and pulmonary function tests. The region of interest of each Chest CT containing the nodule was extracted and analyzed. The radiomic model used the MobileNetV2 extractor followed by a selection of features having positive standard deviation. This filter reduced the 1280 MobileNetV2 features to 333 and the input to an optimized, fully connected network architecture. This architecture had a hidden layer with 150 neurons, dropout with a probability of 0.8, sigmoid activation function, and batch normalization. The clinical model was a logistic regression to predict benignity using 2000 samples generated by bootstrapping to select the best predictors from the clinical variables.
Results
The nodules had a mean diameter of 17.9 mm, histology: 59% adeno, 15% squamous, 2% other malignant, 1% carcinoid, and 23% benign. In the study sample with a malignancy prevalence of 77%, the radiomic model based on visual features exhibited an accuracy of 62% (95% CI: 0.52-0.73) and an AUC of 0.59 (0.49-0.73). The clinical model identified DLCO, tobacco pack-year index, and smoking status as the most consistent clinical predictors associated with the outcome. Integrating the clinical features into the radiomic model resulted in significantly improved accuracy, achieving an accuracy of 79% (95% CI: 69-87%) and an AUC of 0.75 (95% CI: 56-90), increasing 27% in both cases.
Conclusions
Integrating clinical information into a deep learning radiomic model for lung nodule malignancy assessment enhanced the model's accuracy and predictive performance by 27%. This study supports the potential of combined image-based and clinical features to improve lung cancer diagnosis.
Legal entity responsible for the study
Research Institute Germans Trias (IGTP).
Funding
Acadèmia de Ciències Mèdiques de Catalunya i Balears | Barcelona Respiratory Network (BRN)- Fundació Ramon Pla i Armengol Lung Ambition Alliance | Hosptial Germans Trias i Pujol (Llegat JMC) Ministerio de Economía, Industria y Competitividad, Gobierno de España grant number PID 2021-126776OB-C21 Agència de Gestió d'Ajuts Universitaris i de Recerca grant numbers 2021 SGR 01623 and CERCA Programme / Generalitat de Catalunya.
Disclosure
All authors have declared no conflicts of interest.