Scholarlink

<p>Medical image classification plays a vital role in AI-aided medical diagnosis and is often addressed as a multiple instance learning (MIL) issue (i.e., each sample is a bag of instances). For medical images, the disease area or the discriminative area is usually smaller than the whole tissue. In other words, most of the instances in a bag are irrelevant and could interfere with the bag label inference. To address this issue, we add an instance representative selection process before MIL and propose a novel MIL approach named dual space multiple instance representative learning (DSMIRL). DSMIRL consists of two core steps: adaptive instance representative selection (AIRS) and multiple instance representative learning (MIRL). In AIRS, we group and score instances, and then meticulously devise group-wise or instance-wise strategies to select the final collection of instance representatives. The group-wise approaches only preserve the group of instances with the highest instance score as the instance representatives, while the instance-wise ones select the top-<i>k</i> ranked instances in each group to yield the final instance representation collection. In MIRL, we perform aggregations on the selected instance representatives. These aggregations are carried out in both label and feature spaces, so as to further exploit the complementary information of these two spaces. It is worthwhile to point out that this MIRL step can be also flexibly replaced by other existing MIL approaches and enables further improving them. Extensive experiments on four medical image datasets demonstrate the promising performance of DSMIRL over the state-of-the-art MIL approaches and also validate the effectiveness of boosting other MIL by DSMIRL.</p>

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

Adaptive learning of instance representatives in dual spaces for medical image classification

  • 期刊论文

  • 发表日期：

    2025年7月31日

查看全文
  • Xiang Zhu,
  • Sheng Huang,
  • Wenhao Tang,
  • Yi Zhang,
  • Xiaoxian Zhang,
  • Chen Liu,
  • Xiahong Zhang

摘要

Medical image classification plays a vital role in AI-aided medical diagnosis and is often addressed as a multiple instance learning (MIL) issue (i.e., each sample is a bag of instances). For medical images, the disease area or the discriminative area is usually smaller than the whole tissue. In other words, most of the instances in a bag are irrelevant and could interfere with the bag label inference. To address this issue, we add an instance representative selection process before MIL and propose a novel MIL approach named dual space multiple instance representative learning (DSMIRL). DSMIRL consists of two core steps: adaptive instance representative selection (AIRS) and multiple instance representative learning (MIRL). In AIRS, we group and score instances, and then meticulously devise group-wise or instance-wise strategies to select the final collection of instance representatives. The group-wise approaches only preserve the group of instances with the highest instance score as the instance representatives, while the instance-wise ones select the top-k ranked instances in each group to yield the final instance representation collection. In MIRL, we perform aggregations on the selected instance representatives. These aggregations are carried out in both label and feature spaces, so as to further exploit the complementary information of these two spaces. It is worthwhile to point out that this MIRL step can be also flexibly replaced by other existing MIL approaches and enables further improving them. Extensive experiments on four medical image datasets demonstrate the promising performance of DSMIRL over the state-of-the-art MIL approaches and also validate the effectiveness of boosting other MIL by DSMIRL.