Scholarlink

Visual Question Answering (VQA) models have become essential in medical imaging because they enhance patient engagement and aid in clinical decisions. This study introduces a VQA model combining VGG16 for visual feature extraction, Word2Vec for question tokenization, and LSTM for natural language understanding to process fundus images and answer questions regarding diabetic macular edema (DME) grading. The model achieved a training accuracy of 96.88% and a validation accuracy of 87.52%, outperforming ResNet101 in identifying key dataset features. Integrating Word2Vec and LSTM enables the accurate comprehension of complex medical queries. Notably, the model consistently answered related questions without explicit rules. Its simple multimodal fusion approach improves interpretability and computational efficiency without sacrificing performance. Future research could incorporate formal consistency metrics and advanced architectures such as transformers, attention mechanisms, and external knowledge sources to enhance logical coherence, balance accuracy, and consistency. These results highlight the potential of the proposed VQA model to provide reliable and interpretable outcomes in clinical settings and promote its broader use in medical imaging.

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

Vision-Aided Intelligence with Visual Question Answering for Medical Imaging

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Khushi Chalageri,
  • Pragatilaxmi Itigowni,
  • Disha Kalyanshettar,
  • Saakshi Lokhande,
  • Channabasappa Muttal,
  • Vaishnavi J. Ajjevadeyarmath

摘要

Visual Question Answering (VQA) models have become essential in medical imaging because they enhance patient engagement and aid in clinical decisions. This study introduces a VQA model combining VGG16 for visual feature extraction, Word2Vec for question tokenization, and LSTM for natural language understanding to process fundus images and answer questions regarding diabetic macular edema (DME) grading. The model achieved a training accuracy of 96.88% and a validation accuracy of 87.52%, outperforming ResNet101 in identifying key dataset features. Integrating Word2Vec and LSTM enables the accurate comprehension of complex medical queries. Notably, the model consistently answered related questions without explicit rules. Its simple multimodal fusion approach improves interpretability and computational efficiency without sacrificing performance. Future research could incorporate formal consistency metrics and advanced architectures such as transformers, attention mechanisms, and external knowledge sources to enhance logical coherence, balance accuracy, and consistency. These results highlight the potential of the proposed VQA model to provide reliable and interpretable outcomes in clinical settings and promote its broader use in medical imaging.