Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Sam-Guided Enhanced Fine-Grained Encoding with Mixed Semantic Learning for Medical Image Captioning

Zhenyu Zhang, Benlu Wang, Weijie Liang, Yizhi Li, Xuechen Guo, Guanhong Wang, Shiyan Li, Gaoang Wang

TL;DR

Medical image captioning requires detailed region-level descriptions that general multimodal pretraining often fails to provide. The paper introduces MSMedCap, a SAM-guided dual-encoder architecture that combines a CLIP-based encoder for global semantics with a SAM-based encoder for fine-grained details, connected via dual Q-Formers to an OPT LLM. A mixed semantic pre-training regime preserves both broad and fine-grained image-text alignments using ITM, ITG, and ITC objectives, with encoders frozen during captioning and only the Q-Formers and projection layers trained. Experiments on ROCO and MedICaT show MSMedCap outperforms BLIP2 and SAM-BLIP2 across standard metrics, validating the benefit of combining SAM-guided detail capture with general semantic knowledge for medically accurate captions. This approach offers a practical path to enhanced diagnostic captioning by marrying global context with precise anatomical or pathological details.

Abstract

With the development of multimodality and large language models, the deep learning-based technique for medical image captioning holds the potential to offer valuable diagnostic recommendations. However, current generic text and image pre-trained models do not yield satisfactory results when it comes to describing intricate details within medical images. In this paper, we present a novel medical image captioning method guided by the segment anything model (SAM) to enable enhanced encoding with both general and detailed feature extraction. In addition, our approach employs a distinctive pre-training strategy with mixed semantic learning to simultaneously capture both the overall information and finer details within medical images. We demonstrate the effectiveness of this approach, as it outperforms the pre-trained BLIP2 model on various evaluation metrics for generating descriptions of medical images.

Sam-Guided Enhanced Fine-Grained Encoding with Mixed Semantic Learning for Medical Image Captioning

TL;DR

Medical image captioning requires detailed region-level descriptions that general multimodal pretraining often fails to provide. The paper introduces MSMedCap, a SAM-guided dual-encoder architecture that combines a CLIP-based encoder for global semantics with a SAM-based encoder for fine-grained details, connected via dual Q-Formers to an OPT LLM. A mixed semantic pre-training regime preserves both broad and fine-grained image-text alignments using ITM, ITG, and ITC objectives, with encoders frozen during captioning and only the Q-Formers and projection layers trained. Experiments on ROCO and MedICaT show MSMedCap outperforms BLIP2 and SAM-BLIP2 across standard metrics, validating the benefit of combining SAM-guided detail capture with general semantic knowledge for medically accurate captions. This approach offers a practical path to enhanced diagnostic captioning by marrying global context with precise anatomical or pathological details.

Abstract

With the development of multimodality and large language models, the deep learning-based technique for medical image captioning holds the potential to offer valuable diagnostic recommendations. However, current generic text and image pre-trained models do not yield satisfactory results when it comes to describing intricate details within medical images. In this paper, we present a novel medical image captioning method guided by the segment anything model (SAM) to enable enhanced encoding with both general and detailed feature extraction. In addition, our approach employs a distinctive pre-training strategy with mixed semantic learning to simultaneously capture both the overall information and finer details within medical images. We demonstrate the effectiveness of this approach, as it outperforms the pre-trained BLIP2 model on various evaluation metrics for generating descriptions of medical images.
Paper Structure (9 sections, 6 equations, 4 figures, 1 table)

This paper contains 9 sections, 6 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Method
  3. Model Architecture
  4. Mixed Semantic Pre-training
  5. Captioning with the Frozen LLM
  6. Experiments and Results
  7. Experimental Setup
  8. Experimental Design and Results
  9. Conclusion

Figures (4)

  • Figure 1: Overview of (a) BLIP2 and (b) MSMedCap (Ours). MSMedCap implements a supplemented fine-grained feature extraction to synergize functionalities between encoders.
  • Figure 2: Model Architecture. Utilizing frozen dual image encoders and trainable dual Q-Formers enables the extraction of image features at different granularities. During the full model training, only the parameters of the Q-Formers, Soft Queries, and Linear Layers can be trained.
  • Figure 3: Mixed Semantic Pre-training. For the CLIP Q-Former, we train it using the general datasets. As for the SAM Q-Former, we simultaneously train it on both the general and medical datasets.
  • Figure 4: Results Visualization. In our experimental results, we selected several sets of images for visual presentation, comparing our model (MSMedCap) with the baseline (BLIP2 li2023blip2) and the ground truth (GT).