TCM-Net: A Tri-Branch Cross-Modal Network for Intelligent Screening of Pediatric Leukemia
Huazhong Chen, Xuechen Li, Ke Cao, Weibin Zhang, Zhiqiang Liao, Li-Ting Huang · 2025
In recent years, multi-modal fusion technology driven by deep learning has achieved significant progress in medical diagnostics. However, current studies on pediatric acute leukemia screening often focus on uni-modal data, overlooking potential synergies between various CBC sources. This paper proposes a novel tri-modal fusion model-Tri-Branch CrossModal Network (TCM-Net)-to enhance screening accuracy of immature cells in peripheral blood. TCM-Net integrates scattergrams (DIFF and WNB) and MMPB-ALL table data. The model combines a dual-branch ConvNeXtV2 feature extractor, a Cross-Modal-Transformer fusion module, and a joint classifier. Experiments on a real-world dataset of 15,537 clinical samples demonstrate the superior performance of TCM-Net, achieving 86.80% accuracy and an AUC of 0.95, outperforming multiple unimodal and fusion baselines. Ablation studies further verify the effectiveness of the fusion strategy and Cross-Modal-Transformer. This research provides a feasible solution for intelligent leukemia screening with multi-modal medical data.