Scholarlink

<p>This study investigates the effectiveness of EfficientNet as a backbone network for medical image caption generation with the goal of improving the accuracy and descriptive quality of automatically generated medical captions. We propose a deep learning framework that utilizes EfficientNet’s pre-trained visual features, followed by a custom top model consisting of average pooling, dense layers, and dropout for regularization, which interfaces with a Transformer-based decoder for sequence generation. The model is trained using a cross-entropy loss function and optimized with the Adam optimizer. Experiments are conducted on the Radiography Captions (RGC) dataset and performance is compared against the FastVisionModel (Llama-3.2-11B-Vision). The proposed approach achieves an accuracy of 85.0%, a precision of 55.0%, and a cross-entropy loss of 0.50. In addition, standard sequence evaluation metrics demonstrate strong captioning performance, with a BLEU score of 0.453, a ROUGE-L score of 0.805, and a METEOR score of 0.731. These results indicate that the proposed EfficientNet–Transformer architecture effectively captures both visual and semantic information in complex medical images, producing accurate and informative captions and highlighting its potential for enhancing interpretability in medical imaging and supporting clinical decision-making systems.</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

Innovative deep learning model for medical image captioning using EfficientNet and transformer

  • 期刊论文

  • 发表日期：

    2026年6月3日

查看全文
  • Salma Elgayar,
  • Ibrahim I. M. Manhrawy,
  • Shuang Wang,
  • Safwat Hamad,
  • El-Sayed M. Horbaty

摘要

This study investigates the effectiveness of EfficientNet as a backbone network for medical image caption generation with the goal of improving the accuracy and descriptive quality of automatically generated medical captions. We propose a deep learning framework that utilizes EfficientNet’s pre-trained visual features, followed by a custom top model consisting of average pooling, dense layers, and dropout for regularization, which interfaces with a Transformer-based decoder for sequence generation. The model is trained using a cross-entropy loss function and optimized with the Adam optimizer. Experiments are conducted on the Radiography Captions (RGC) dataset and performance is compared against the FastVisionModel (Llama-3.2-11B-Vision). The proposed approach achieves an accuracy of 85.0%, a precision of 55.0%, and a cross-entropy loss of 0.50. In addition, standard sequence evaluation metrics demonstrate strong captioning performance, with a BLEU score of 0.453, a ROUGE-L score of 0.805, and a METEOR score of 0.731. These results indicate that the proposed EfficientNet–Transformer architecture effectively captures both visual and semantic information in complex medical images, producing accurate and informative captions and highlighting its potential for enhancing interpretability in medical imaging and supporting clinical decision-making systems.