A1310
Title: Image-based whole-transcriptome deep learning predicts prognosis in lung adenocarcinoma
Authors: Yang-Ming Yeh - Academia Sinica (Taiwan)
Yahsuan Chang - National Health Research Institutes (Taiwan)
Hsuan-Yu Chen - Academia Sinica (Taiwan)
Hsuanyu Chen - Academia Sinica (Taiwan) [presenting]
Abstract: Targeted therapies and immunotherapy have improved outcomes in lung adenocarcinoma, yet recurrence and limited survival remain major challenges. An image-based deep learning model for prognosis prediction using whole-transcriptome data was developed and evaluated across four cohorts including one RNA-seq training cohort (n=391) and three independent validation cohorts (TCGA, n=394; GSE68465, n=443; GSE13213, n=117). Gene expression data were transformed into images via pixel encoding. A convolutional neural network was trained using all genes without feature selection. The training cohort was split into training (n=259), testing (n=62), and validation (n=70) subsets, followed by external validation. The model effectively stratified patients into high- and low-risk groups. High-risk patients showed significantly worse overall survival in the training cohort (P<0.01) and in TCGA (P=0.001), GSE68465 (P<0.0001), and GSE13213 (P=0.003). Prediction accuracies were 0.71, 0.81, and 0.92 in the training, testing, and validation subsets, respectively, and 0.70 across external cohorts. This image-based deep learning framework enables robust, platform-independent prognostic prediction without gene selection, supporting scalable precision oncology.
