CARZero: Cross-Attention Alignment for Radiology Zero-Shot Classification

Abstract: The advancement of Zero-Shot Learning in the medical domain has been driven forward by using pre-trained models on large-scale image-text pairs, focusing on image-text alignment. However, existing methods primarily rely on cosine similarity for alignment, which may not fully capture the complex relationship between medical images and reports. To address this gap, we introduce a novel approach called Cross-Attention Alignment for Radiology Zero-Shot Classification (CARZero). Our approach innovatively leverages cross-attention mechanisms to process image and report features, creating a Similarity Representation that more accurately reflects the intricate relationships in medical semantics. This representation is then linearly projected to form an image-text similarity matrix for cross-modality alignment. Additionally, recognizing the pivotal role of prompt selection in zero-shot learning, CARZero incorporates a LLM-based prompt alignment strategy. This strategy standardizes diverse diagnostic expressions into a unified format for both training and inference phases, overcoming the challenges of manual prompt design. Our approach is simple yet effective, demonstrating state-of-the-art performance in zero-shot classification on five official chest radiograph diagnostic test sets, including remarkable results on datasets with long-tail distributions of rare diseases. This achievement is attributed to our new image-text alignment strategy, which effectively addresses the complex relationship between medical images and reports. Code and models are available at https://github.com/laihaoran/CARZero.

• The paper introduces a novel cross-attention alignment mechanism combined with LLM-based prompt strategies to improve zero-shot radiology classification.
• It demonstrates state-of-the-art AUC performance on multiple chest radiograph datasets, notably excelling in identifying rare diseases and long-tail distributions.
• The approach reduces dependency on labeled data while setting a new benchmark for multi-modal diagnostic systems in medical imaging.

CARZero: Cross-Attention Alignment for Radiology Zero-Shot Classification

Introduction

The rapid progress in deep learning (DL) has catalyzed advancements in medical image recognition, primarily focusing on tasks such as disease diagnosis. Despite impressive results, these systems predominantly depend on labeled data, which are often scarce and expensive to obtain in a clinical setting. Zero-Shot Learning (ZSL), leveraging pre-trained models on extensive image-text datasets, has emerged as a promising avenue, particularly advantageous for diagnosing rare diseases. However, the standard practices primarily employ cosine similarity for image-text alignment, which fails to encapsulate the intricate correlations present between medical imagery and corresponding textual reports.

The paper "CARZero: Cross-Attention Alignment for Radiology Zero-Shot Classification" offers a novel framework termed CARZero (Cross-Attention Radiology Zero-Shot), which addresses the complexity of medical image-text relationships via a cross-attention mechanism. This approach redefines the paradigm of zero-shot medical diagnostics by introducing a more nuanced similarity representation and leveraging LLM-based prompt alignment to optimize diagnosis.

Methods

CARZero utilizes a two-stage architecture that integrates cross-attention alignment and an LLM-based strategy for prompt alignment. The primary objective is to establish a robust alignment between image and text features using cross-attention, thus overcoming the limitations of simplistic cosine similarity measures.

Cross-Attention Alignment: As depicted (Figure 1), the proposed model introduces a cross-attention module to generate a high-dimensional Similarity Representation (SimR). This SimR acts as a descriptor of the complex semantic relationships between images and corresponding radiology reports. It involves global and local feature extraction from both modalities, producing features that capture mutual interactions, which are subsequently optimized using InfoNCE loss.

Figure 1: Comparison of the alignment scheme in Visual Language Pre-training: (left) handcrafted cosine similarity used in CLIP~\cite{conde2021clip}.

LLM-based Prompt Alignment: Recognizing the challenging nature of prompt crafting, especially in medical contexts, CARZero integrates LLMs to convert varying diagnostic expressions into standardized prompts, thereby mitigating manual prompt design efforts. This aligns the training and testing phases more effectively, enhancing zero-shot learning performance.

Figure 2: The \algname Network proposed in this paper consists of two stages. First, LLM is employed to generate prompt templates from medical reports. Second, text and vision encoders are used to extract features from image and text, which are fed into a cross-attention module to generate similarity for optimizing InfoNCE loss.

Experimental Evaluation

Extensive experiments conducted on five publicly available chest radiograph datasets underscore the efficacy of the CARZero framework. The method achieves state-of-the-art AUC performance particularly noted on datasets characterized by long-tail distributions such as PadChest (0.810). Furthermore, CARZero also excels on datasets focusing on rare diseases, demonstrating its potential for practical diagnostic applications.

Comparison with Existing Methods: When evaluated against existing state-of-the-art approaches, CARZero consistently outperforms in zero-shot settings across multiple datasets, highlighting its superior alignment and generalization capabilities.

Figure 3: Visualization of attention map in \algname on ChestXDet10. The red boxes indicate the corresponding ground truth of detection. Highlighted pixels represent higher activation weights correlating specific words with regions in the image.

Implications and Future Directions

The introduction of CARZero represents a significant enhancement in zero-shot medical classification, specifically by effectively aligning complex medical image-text modalities. This approach not only establishes a more accurate diagnostic tool but also reduces reliance on extensive labeled datasets, potentially revolutionizing the way rare diseases are diagnosed and understood.

Looking forward, CARZero's cross-attention alignment strategy could inspire further research in both fine-tuning applications and its extension to natural image domains. Moreover, enhancing interpretability and exploring multi-modal pre-training across other types of medical imaging remain promising avenues for future exploration.

Conclusion

CARZero offers a valuable contribution to zero-shot classification by engineering a cross-attention alignment mechanism that captures intricate interactions between medical images and textual reports. Its innovative use of LLMs for prompt alignment further solidifies its performance, setting a benchmark for future systems in clinical AI diagnostics. With its advanced capabilities in zero-shot classification, CARZero not only demonstrates practical applicability in real-world scenarios but also illuminates the path toward more inclusive and robust diagnostic frameworks in medical AI.