CAMANet: Class Activation Map Guided Attention Network for Radiology Report Generation

TL;DR

A Class Activation Map guided Attention Network (CAMANet) is proposed which explicitly promotes cross-modal alignment by employing aggregated class activation maps to supervise cross-modal attention learning, and simultaneously enrich the discriminative information.

Abstract

Radiology report generation (RRG) has gained increasing research attention because of its huge potential to mitigate medical resource shortages and aid the process of disease decision making by radiologists. Recent advancements in RRG are largely driven by improving a model's capabilities in encoding single-modal feature representations, while few studies explicitly explore the cross-modal alignment between image regions and words. Radiologists typically focus first on abnormal image regions before composing the corresponding text descriptions, thus cross-modal alignment is of great importance to learn a RRG model which is aware of abnormalities in the image. Motivated by this, we propose a Class Activation Map guided Attention Network (CAMANet) which explicitly promotes crossmodal alignment by employing aggregated class activation maps to supervise cross-modal attention learning, and simultaneously enrich the discriminative information. CAMANet contains three complementary modules: a Visual Discriminative Map Generation module to generate the importance/contribution of each visual token; Visual Discriminative Map Assisted Encoder to learn the discriminative representation and enrich the discriminative information; and a Visual Textual Attention Consistency module to ensure the attention consistency between the visual and textual tokens, to achieve the cross-modal alignment. Experimental results demonstrate that CAMANet outperforms previous SOTA methods on two commonly used RRG benchmarks.

Figures (8)

• Figure 1: A Chest XRay image with its report findings. By way of example, manually aligned visual-textual features are marked in the same color
• Figure 2: The architecture of CAMANet: An image is fed into the Visual Extractor to obtain patch features which are then utilized to generate the VDM via the VDM Generation module. The proposed VDMAE leverages the VDM to derive a discriminative representation and enrich the discriminative information. Combined with word embedding, visual tokens are sent to the transformer to produce the report. After that, the VTAC module generates a TDM from the cross-modal attention scores and considers the VDM as the ground truth to supervise the cross-modal alignment learning.
• Figure 3: An illustration of the visual-textual attention consistency. The decoder model is expected to pay attention to the same regions as the vision model to achieve consistency.
• Figure 4: Effect of varying $k$ on MIMIC-CXR. (BLEU-4 score).
• Figure 5: Effect of varying $k$ on IU-Xray. (BLEU-4 score).