<p>Visual diagnosis training in medicine often struggles to efficiently build learners’ skill, with existing instructional approaches lacking consensus and showing variable effectiveness. This study introduces a two-phase conceptual framework for visual diagnosis training, using electrocardiogram interpretation as an example to explore how novices develop discriminatory skills and diagnostic certainty. Through a prospective cohort study involving 78 medical students and residents who completed at least 75 cases with demonstrated improvement, we applied a version of item-response modeling to analyze their learning trajectories. Findings showed that diagnostic cases could be represented on a continuous scale from normal to abnormal, with a “confusability zone” where cases are difficult to classify. Participants’ ability to calibrate diagnostic certainty improved modestly, reflected in more organized use of certainty ratings. Learning curves showed that the ability to distinguish normal from abnormal developed earlier and more rapidly than the capacity to assign specific diagnostic labels, with both skills improving over time but at different rates. These results suggest that visual diagnosis training could be structured in distinct phases, beginning with abnormality detection and progressing to detailed diagnostic classification. The framework supports targeted curriculum design, including the use of confusable cases for focused training and monitoring of diagnostic certainty as an educational assessment tool, with likely broad applicability across medical imaging domains.</p>