Abstract:
Lung cancer, particularly lung adenocarcinoma, is still challenging to forecast using a single source of information. This thesis describes a lightweight multimodal pipeline that combines chest CT scans with RNA-Seq gene expression patterns to improve patient-level prediction. CT volumes are transformed to Hounsfield units, windowed and resampled, and then represented as 2-D slice bags; a compact encoder extracts slice features, which are summarized using attention-based Multiple-Instance Learning (MIL) to build a patient-level CT embedding. To avoid leakage, RNA-Seq is log-standardized and compressed using principal component analysis (PCA) with just training folds fitted. The two modality embeddings are combined (concatenated) and sent to a tiny neural classifier. We evaluated the paired intersection cohort (n≈30, balanced labels) using 5-fold stratified cross-validation. We report both best-per-fold and seed-wise summaries for transparency. The fused model achieved AUROC 0.8400 ± 0.1326, AUPRC 0.8675 ± 0.1033, Accuracy 0.8881 ± 0.0649, and F1 0.8692 ± 0.0825 (best-per-fold), whereas seed-wise CV underlines predicted small- sample variance (AUROC 0.658 ± 0.244; AUPRC 0.759 ± 0.172). MIL focus improves qualitative interpretation by emphasizing a small subset of influential CT slices each patient. Together, the findings confirm the feasibility and utility of CT+RNA fusion, while also encouraging future research into bigger cohorts, external validation, self-supervised CT pretraining, greater MIL and co-attention fusion, and deployment calibration.
