Balancing conflicting demands in medical image report writing

Automatic radiology report generation needs to satisfy two competing demands: match clinical labels precisely and generate fluent natural language. Standard multi-task learning uses simple weighted averaging, which fails because these goals push gradients in opposite directions. The authors diagnose this as a "Double Dilemma" using differential equation theory, then propose CAME-Grad, an optimizer that redirects conflicting gradients and dynamically balances task demands. Tested on eight existing methods across two datasets, it consistently improves clinical metrics by 2–2.3% without changing underlying architectures.