Severity‑aware multimodal network for chest x‑ray triage and disease captioning

Rising radiograph volumes and radiologist workforce shortages create a Triage Bottleneck, where urgent cases queue behind routine ones. This thesis introduces the Cross-Modal Triage Network (CMTN), a multimodal framework fusing a Swin Transformer V2 visual backbone with a PubMedBERT textual encoder via Gated Cross-Attention to prioritize chest radiographs by severity while jointly predicting fourteen pathology labels. Evaluated on 3,475 MIMIC-CXR cases, CMTN achieved strong ordinal agreement with its NLP-derived reference labels (QWK = 0.9341), low calibration error (ECE = 0.0355), and 34 ms end-to-end latency, with native cross-attention heatmaps achieving 54.3% expert-rated localization accuracy. However, a blinded clinical audit revealed substantially lower agreement with genuine radiologist judgment (QWK = 0.1399), indicating that the model reproduces its automated labeling pipeline rather than expert reasoning. This gap identifies the validation work required before clinical deployment. CMTN establishes an efficient, interpretable architecture that is computationally ready for pilot integration, pending comprehensive clinical validation.