115P Spitz tumor classification using artificial intelligence

Abstract

routinely collected clinical metadata, or are trained on historic lower-resolution datasets. Methods:We developed a multimodal AI model separately integrating last-available and longitudinal LDCT imaging with clinical metadata to predict lung nodule malignancy risk.We specifically evaluated challenging nodules of uncertain malignancy probability on single-timepoint imaging.The model was trained using one of the world's largest LCS trials (SUMMIT: recruiting from 2019), which screened >13,000 individuals.A 3D ResNet-18 processed image data.A TabNet encoder processed clinical metadata.Scan interval and clinical variables were incorporated with image data through feature-wise linear modulation.Performance was evaluated with 5-fold cross-validation using the area under the receiver operating characteristic curve (AUC), sensitivity (Se), and specificity (Sp).Results: Across the hold-out test folds, the single-timepoint model achieved an average AUC of 0.790.03(Se: 0.810.07;Sp: 0.660.09).Incorporating longitudinal imaging data improved the AUC to 0.850.03(Se: 0.790.05;Sp: 0.760.03).Adding scan-interval information further increased the AUC to 0.860.02(Se: 0.840.07;Sp: 0.730.09).The best-performing model, combining longitudinal imaging data, scan interval, and clinical metadata, achieved an AUC of 0.870.03(Se: 0.830.09;Sp: 0.800.05).Conclusions: These results emphasise that longitudinal imaging with scan-interval information provides a clear improvement over single-timepoint models for malignancy prediction, especially in challenging cases.The addition of metadata yields modest gains.The consistently high sensitivity across models complements the often-reported high specificity of radiologists.Integrating such tools into radiology workflows may offer complementary diagnostic value in lung cancer screening.

Ähnliche Arbeiten