Algorithm 2 Validation-Tuned Caption-Mediated Late Fusion

Input: reranked candidate list ${\pi }_c^{\left(0\right)}$, case-caption text $g\left(c\right)$, disease caption bank $h\left(d\right)$, candidate weight set $\mathcal{W}$. Output: final ranked list ${\pi }_c$. For each candidate disease $d\in {\pi }_c^{\left(0\right)}$, compute the caption-overlap score $s_{\mathrm{img}}(c,d)$ using Equation (8). Min–max normalize the BM25, classifier, dense-retrieval, reranker, and image scores within case c. For each candidate weight vector $\mathbf{w}\in \mathcal{W}$, compute validation-set rankings with Equation (9). Select the weight vector that maximizes the objective in Equation (10) on the validation split. Apply the selected weights to the validation and test candidates and sort diseases by $s_{\mathrm{final}}(c,d)$.