Align-then-Aggregate: Pretreatment prediction of radiation-induced lymphopenia in nasopharyngeal carcinoma via multi-modal collaborative learning

Pretreatment prediction of radiation-induced lymphopenia (RIL) aims to proactively assess the risk of lymphopenia and improve post-radiotherapy survival, particularly for patients with nasopharyngeal carcinoma. Existing methods are typically based on expert-designed workflows, where handcrafted features are independently extracted from computed tomography (CT) images, dose maps, and clinical factors, and directly combined to predict RIL grades. These methods are subject to human bias, neglect inter-modal heterogeneity and lack effective modeling of complementary inter-modal interaction, which limits the performance of RIL prediction. Based on this, a novel multi-modal collaborative learning framework Align-then-Aggregate Network (AANet) is proposed to achieve accurate RIL prediction in an end-to-end manner using an Align-then-Aggregate strategy, where the alignment stage maps CT and dose features into hierarchically aligned representations that are spatially co-registered and semantically comparable, facilitating the extraction of discriminative patterns to RIL, and the aggregation stage enables modality-specific information exchange guided by patient clinical information, allowing inter-modal complementarity and providing a more comprehensive representation of the patient. Compared to previous methods, AANet can fully automatically extract features from multiple modalities and effectively integrate heterogeneous modalities using a two-stage Align-then-Aggregate strategy and multi-perspective fusion algorithm, thereby improving RIL prediction. Extensive experiments on a newly established dataset demonstrate that AANet achieves accurate RIL prediction with an accuracy of 0.8148±0.0380 and an area under the curve of 0.8562±0.0450, outperforming existing methods.