MoMBS: Mixed-Order Sampling Improves Training on Heterogeneous-Quality Data for Universal Lesion Detection

MCML Authors

Han Li

Dr.

→ Group Nassir Navab
 Computer Aided Medical Procedures & Augmented Reality

Jingsong Liu

→ Group Peter Schüffler
 Computational Pathology

Peter Schüffler

Prof. Dr.

Core PI

Computational Pathology

Abstract

Universal lesion detection (ULD) in computed tomography plays an essential role in computer-aided diagnosis. However, images in ULD tasks often exhibit substantial variations in quality, including issues such as limited clarity and inaccurate labels. Effectively leveraging training images with heterogeneous qualities thus becomes a significant challenge. Existing training strategies, such as self-paced learning (SCL) and hard example mining (OHEM), attempt to address this by reweighting high-loss samples. However, they often rely solely on loss, which inadequately captures sample difficulty, and may lead to over- or under-utilization of hard samples. To tackle this, we revisit the role of minibatch sampling (MBS) and propose a novel Mixed-order Minibatch Sampling (MoMBS) approach. MoMBS introduces a joint measure based on both loss and uncertainty, moving beyond sole reliance on loss. This enables a finer-grained categorization of high-loss samples by distinguishing between those that are poorly labeled and underrepresented, versus those that are overfitted but well represented. Motivated by human learning behavior, we prioritize underrepresented samples as the main contributors to the minibatch gradient and protect them from being overwhelmed by noisy or overfitted examples via a mixed-order sampling design. This leads to a more accurate estimation of sample difficulty and avoids indiscriminate treatment in the utilization of hard examples. Our primary experiments on DeepLesion for ULD show that MoMBS improves over two state-of-the-art (SOTA) methods across three different dataset settings by 0.97%–7.28%. To further verify its generalizability, we evaluate MoMBS on three additional tasks: Seg-19 (up to 2.3% improvement over five SOTA baselines), CIFAR100-LT (up to 5.3% over nine SOTA methods), and CIFAR100-NL (up to 4.6% over five SOTA methods), consistently outperforming existing approaches.

article LLS+26b

---

Medical Image Analysis

In press. Jun. 2026.

---

Authors

H. Li • J. Liu • P. J. Schüffler • H. Han • S. K. Zhou

---

Links

DOI URL

---

Research Area

 C1 | Medicine