Concetti Chiave
Improving the inter-report consistency of radiology report generation by extracting lesions, examining their characteristics, and using a lesion-aware mixup technique to align the representations of semantically equivalent lesions.
Sintesi
The paper proposes ICON, a framework that aims to improve the inter-report consistency of radiology report generation. The key components are:
-
Lesion Extraction (Stage 1):
- ZOOMER: A visual encoder that classifies input images into abnormal observations (lesions) without requiring fine-grained labels (e.g., bounding boxes).
- The extracted lesions are used as input for the report generation stage.
-
Report Generation (Stage 2):
- INSPECTOR: A visual encoder that inspects each lesion and matches it with corresponding attributes to differentiate it from other variations.
- Lesion-Attribute Alignment: A cross-attention module is used to align the lesion representations with the attribute representations.
- Lesion-aware Mixup: A mixup augmentation technique is introduced to ensure that the representations of semantically equivalent lesions align with the same attributes, achieved through a linear combination during the training phase.
The authors conduct extensive experiments on three publicly available chest X-ray datasets (IU X-RAY, MIMIC-CXR, and MIMIC-ABN) and demonstrate that ICON outperforms state-of-the-art baselines in terms of both inter-report consistency and clinical accuracy.
Statistiche
There are small bilateral pleural effusions.
There is no pleural effusion.
There are small bilateral pleural effusions.
There is no pleural effusion.
There are small bilateral pleural effusions.
There is no pleural effusion.
Citazioni
"To the best of our knowledge, we are the first to introduce inter-report consistency in radiology report generation."
"ICON only requires coarse-grained labels (i.e., image labels) for training to extract lesions, in contrast to previous methods that require fine-grained labels (i.e., bounding boxes)."
"Extensive experiments are conducted on three publicly available datasets, and the results demonstrate the effectiveness of ICON in terms of improving both the consistency and accuracy of the generated reports."