정밀의료를 위한 멀티모달 딥러닝 기반 ADPKD 분류 모델

초록

Autosomal Dominant Polycystic Kidney Disease (ADPKD) is a common genetic disorder leading to kidney failure, affecting millions globally. This study presents a non-invasive approach to diagnose genetic mutations using T2-weighted MR images and patient clinical data, advancing precision medicine. We analyzed 441 ADPKD cases from multiple institutions, categorized into PKD1 Truncating (n=232) and other PKD mutations (n=209). MR images were preprocessed for uniformity, including masking, resampling to a 0.78×0.78×0.78 mm³ voxel size, normalization, and resizing to (256, 256, 64). Three models were evaluated using accuracy, F1-score, and AUC: a Unimodal Model with MR images only, a Unimodal Model with MR images and masks, and a Multimodal Model integrating MR images, masks, and patient clinical data (age and height). The Multimodal Model, employing a 3D ResNet-152 and Multilayer Perceptron (MLP), achieved the highest performance with 73.9% accuracy and a 0.754 F1-score. These results demonstrate the effectiveness of integrating multiple data modalities for enhanced classification of ADPKD genetic mutations, offering a promising tool for non-invasive diagnosis and improved clinical decision-making.

키워드