Scholarlink

<p>Correctly classifying brain tumors is imperative to the prompt and accurate treatment of a patient. In this work, we first study the difficult but realistic setting of training a deep learning brain tumor classification model in the presence of noisy MR images. Then, we consider the emerging setting of vision-language models (VLMs), and more specifically study how VLM in-context learning can be leveraged to solve the brain tumor classification task. For both settings, we propose methods that are inspired by influence functions which stem from the field of robust statistics. For the deep learning setting, we propose two training methods that are robust to noisy MRI training data: influence-based sample reweighing (ISR) and influence-based sample perturbation (ISP). In ISR, we adaptively reweigh training examples according to how helpful/harmful they are to the training process, while in ISP, we craft and inject “healthy” noise proportional to the influence score. For the VLM setting, we propose a demonstration selection method called influence-based in-context learning (IICL). With IICL, influence score is leveraged to find the most suitable MRI samples that can demonstrate the assigned task to the pretrained VLMs. Both ISR and ISP harden the classification model against noisy training data without significantly affecting the generalization ability of the model on test data, while IICL significantly boosts the performance of GPT-4o and Gemini. We conduct empirical evaluations over a common brain tumor dataset and compare our methods against several baselines to demonstrate the advantage of our three methods.</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

Influence-based approaches for tumor classification in noisy brain MRI with deep learning and vision-language models

  • 期刊论文

  • 发表日期：

    2025年6月17日

查看全文
  • Minh-Hao Van,
  • Alycia N. Carey,
  • Xintao Wu

摘要

Correctly classifying brain tumors is imperative to the prompt and accurate treatment of a patient. In this work, we first study the difficult but realistic setting of training a deep learning brain tumor classification model in the presence of noisy MR images. Then, we consider the emerging setting of vision-language models (VLMs), and more specifically study how VLM in-context learning can be leveraged to solve the brain tumor classification task. For both settings, we propose methods that are inspired by influence functions which stem from the field of robust statistics. For the deep learning setting, we propose two training methods that are robust to noisy MRI training data: influence-based sample reweighing (ISR) and influence-based sample perturbation (ISP). In ISR, we adaptively reweigh training examples according to how helpful/harmful they are to the training process, while in ISP, we craft and inject “healthy” noise proportional to the influence score. For the VLM setting, we propose a demonstration selection method called influence-based in-context learning (IICL). With IICL, influence score is leveraged to find the most suitable MRI samples that can demonstrate the assigned task to the pretrained VLMs. Both ISR and ISP harden the classification model against noisy training data without significantly affecting the generalization ability of the model on test data, while IICL significantly boosts the performance of GPT-4o and Gemini. We conduct empirical evaluations over a common brain tumor dataset and compare our methods against several baselines to demonstrate the advantage of our three methods.