Scholarlink

Low Back Pain diagnosis relies heavily on the precise determination of clinical data, currently, many existing datasets suffer from severe multi-label class imbalance. This imbalanced data causes standard models to prioritize majority classes, resulting in critically low recall for rare but severe pathologies. In this paper, we propose a novel diagnostic framework based on MedCLIP to mitigate this imbalance. We systematically evaluate three distinct strategies: (1) Over-Sampling with Label Powerset (OSLP), (2) a novel proposed CLIP-adapted Deferred Re-Weighting (DRW) scheme, and (3) a Hybrid approach integrating DRW with Mixup augmentation. Extensive experiments demonstrate that while OSLP improves decision boundaries, it yields only marginal sensitivity gains. Conversely, our Hybrid approach significantly outperforms the baseline, boosting Recall from 0.4834 to 0.6814 (an improvement of approximately 41%) while achieving the highest AUPRC of 0.7259. Our results highlight the efficiency of combining algorithmic re-weighting with data-oriented regularization to ensure robust, high-sensitivity medical diagnosis.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Novel Adaptation of a MedCLIP-Based Vision-Language Model for Imbalanced Multi-Label Datasets

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Tran Tan Ly,
  • Quoc Binh Le,
  • Thu T. B. Le,
  • Quang Tran Minh,
  • Thi Cuc Le,
  • Tan Ha Mai,
  • Trong Nhan Phan

摘要

Low Back Pain diagnosis relies heavily on the precise determination of clinical data, currently, many existing datasets suffer from severe multi-label class imbalance. This imbalanced data causes standard models to prioritize majority classes, resulting in critically low recall for rare but severe pathologies. In this paper, we propose a novel diagnostic framework based on MedCLIP to mitigate this imbalance. We systematically evaluate three distinct strategies: (1) Over-Sampling with Label Powerset (OSLP), (2) a novel proposed CLIP-adapted Deferred Re-Weighting (DRW) scheme, and (3) a Hybrid approach integrating DRW with Mixup augmentation. Extensive experiments demonstrate that while OSLP improves decision boundaries, it yields only marginal sensitivity gains. Conversely, our Hybrid approach significantly outperforms the baseline, boosting Recall from 0.4834 to 0.6814 (an improvement of approximately 41%) while achieving the highest AUPRC of 0.7259. Our results highlight the efficiency of combining algorithmic re-weighting with data-oriented regularization to ensure robust, high-sensitivity medical diagnosis.