Class imbalance in medical imaging datasets remains a key challenge for the classification stages of object detection pipelines, particularly when rare yet clinically significant pathologies coexist with prevalent findings. In spinal MRI, common conditions such as Normal Intervertebral Disc (IVD) may constitute over 45% of annotated objects, whereas findings like Spondylolisthesis account for fewer than 2% of instances. Conventional loss functions including Focal Loss, Class-Balanced Loss, and Label-Distribution-Aware Margin Loss, each address individual aspects of this imbalance but do not provide a unified, adaptive solution. Inspired by ensemble loss strategies developed in Deep Metric Learning (DML), we propose WELDE (Weighted Ensemble Loss with Diversity Enhancement), a framework that combines four complementary loss functions via per-head adapter projections, EMA-based normalization, and learnable adaptive weighting with a relaxed sum-to-one penalty. Each loss component receives a dedicated classification head with an independent adapter projection from a shared frozen backbone, enabling feature specialization without backbone fine-tuning. We provide theoretical analysis of WELDE’s properties, including gradient magnitude balancing across loss components and weight non-degeneracy. Applied to a lumbar mid-sagittal spinal MRI dataset with six classes and a 33.9:1 imbalance ratio, WELDE demonstrates consistent, though incremental, improvements in classification performance among the evaluated methods. It outperforms all single-loss baselines (mAP 0.702 vs. 0.689 for the best baseline CE, mAP<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(_{\text {tail}}\)</EquationSource> </InlineEquation> 0.509 vs. 0.472, +8.1% relative improvement on tail classes) and an architecture-matched CE ensemble control (mAP<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(_{\text {tail}}\)</EquationSource> </InlineEquation> 0.509 vs. 0.496), indicating that the empirical improvement derives from diverse loss composition rather than increased model capacity. External cross-domain validation on the DermaMNIST skin lesion benchmark (7 classes, <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\rho =58.3\)</EquationSource> </InlineEquation>) confirms that WELDE generalizes well, achieving the highest mAP (0.709) and mAP<InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(_{\text {tail}}\)</EquationSource> </InlineEquation> (0.651) among all methods, placing above both single-head baselines (<InlineEquation ID="IEq5"> <EquationSource Format="TEX">\(+11.5\%\)</EquationSource> </InlineEquation> mAP over CE) and the architecture-matched CE ensemble control.

WELDE: a weighted ensemble loss with diversity enhancement for imbalanced object classification in medical imaging

期刊论文
发表日期：

2026年5月30日

查看全文

Rao Farhat Masood,
Imtiaz Ahmad Taj

摘要

Class imbalance in medical imaging datasets remains a key challenge for the classification stages of object detection pipelines, particularly when rare yet clinically significant pathologies coexist with prevalent findings. In spinal MRI, common conditions such as Normal Intervertebral Disc (IVD) may constitute over 45% of annotated objects, whereas findings like Spondylolisthesis account for fewer than 2% of instances. Conventional loss functions including Focal Loss, Class-Balanced Loss, and Label-Distribution-Aware Margin Loss, each address individual aspects of this imbalance but do not provide a unified, adaptive solution. Inspired by ensemble loss strategies developed in Deep Metric Learning (DML), we propose WELDE (Weighted Ensemble Loss with Diversity Enhancement), a framework that combines four complementary loss functions via per-head adapter projections, EMA-based normalization, and learnable adaptive weighting with a relaxed sum-to-one penalty. Each loss component receives a dedicated classification head with an independent adapter projection from a shared frozen backbone, enabling feature specialization without backbone fine-tuning. We provide theoretical analysis of WELDE’s properties, including gradient magnitude balancing across loss components and weight non-degeneracy. Applied to a lumbar mid-sagittal spinal MRI dataset with six classes and a 33.9:1 imbalance ratio, WELDE demonstrates consistent, though incremental, improvements in classification performance among the evaluated methods. It outperforms all single-loss baselines (mAP 0.702 vs. 0.689 for the best baseline CE, mAP \(_{\text {tail}}\) 0.509 vs. 0.472, +8.1% relative improvement on tail classes) and an architecture-matched CE ensemble control (mAP \(_{\text {tail}}\) 0.509 vs. 0.496), indicating that the empirical improvement derives from diverse loss composition rather than increased model capacity. External cross-domain validation on the DermaMNIST skin lesion benchmark (7 classes, \(\rho =58.3\) ) confirms that WELDE generalizes well, achieving the highest mAP (0.709) and mAP \(_{\text {tail}}\) (0.651) among all methods, placing above both single-head baselines ( \(+11.5\%\) mAP over CE) and the architecture-matched CE ensemble control.