AI for Chest X-Ray Analysis: Classifying Six Pathologies with RAD-DINO-MAIRA-2

From October 2025 to April 2026, the author built a research prototype that takes a chest X-ray and outputs six independent probabilities (0-1) for pathologies: cardiomegaly, atelectasis, consolidation, pleural effusion, pneumothorax, and edema. The approach uses a frozen Microsoft RAD-DINO-MAIRA-2 encoder (DINOv2 pretrained on ~1.4 million X-rays) with a small head (LayerNorm → Linear(768→256) → GELU → Dropout → Linear(256→6)), training only 199,942 of 86.8 million parameters. Training used AdamW, batch size 64, bf16 precision, and cost about $300 on an A100 80GB on RunPod. Data came from VinBigData (15,000 images) and OpenI (7,466 images), with OpenI labels extracted via rule-based parsing of radiology reports, totaling 22,466 images split 85/15. The author attempted DenseNet121 first but abandoned it due to failure to focus on relevant anatomy; heatmaps showed attention on artifacts and irrelevant regions. The project explicitly avoids segmentation or report generation and is not certified for clinical use; the author seeks an organization for clinical trials to register as a Class 3 medical device in Russia.