TedMIL: Text-Driven Multiple Instance Learning for Fine-Grained Pathological Classification with Melanoma-FISH Status Prediction

Fine-grained classification of melanoma with fluorescence in situ hybridization (FISH) status prediction using whole slide images (WSIs) provides critical insights into progression and prognosis. Multiple instance learning (MIL) has been effectively applied. However, existing MIL methods rely on visual features, which makes MIL lack macroscopic prior-driven guidance, resulting in the inability of visual modeling to better capture the intrinsic representations of melanoma. To fully model the macroscopic priors provided by category-agnostic texts, this study proposes TedMIL, a novel Text-driven MIL, which consists of the Text Injected Re-embedding (TIR) and the Text-Image Matching (TIM). Based on pathological priors, LLMs, and gated feature fusion, TIR re-embeds the feature representations of MIL by injecting macroscopic sementic into morphological features. TIM refines the final class output probabilities through the cross-modal affinity quantification mechanism based on normalized feature space distance. Data were accessed for research purposes between 25/10/2025 and 24/10/2026 in accordance with the approved study protocol. For the melanoma identification task, TedMIL achieved an AUC of 97.9% and an accuracy (ACC) of 93.4% on the test set, demonstrating impressive discriminative capability. For the more challenging fine-grained three-class classification task, TedMIL improves accuracy (ACC) by 2.8%, AUC by 2.2%, and F1-score by 3.0% compared with state-of-the-art MIL baselines. Qualitative visualizations further show that text guidance leads to more spatially coherent attention patterns that better align with diagnostically relevant morphological regions. This paper demonstrates that the proposed text-driven MIL could realize more intrinsic pathological representations and achieve efficient fine-grained classification of melanoma.