Diagnose like a Radiologist: Attention Guided Convolutional Neural Network for Thorax Disease Classification (1801.09927v1)

Abstract: This paper considers the task of thorax disease classification on chest X-ray images. Existing methods generally use the global image as input for network learning. Such a strategy is limited in two aspects. 1) A thorax disease usually happens in (small) localized areas which are disease specific. Training CNNs using global image may be affected by the (excessive) irrelevant noisy areas. 2) Due to the poor alignment of some CXR images, the existence of irregular borders hinders the network performance. In this paper, we address the above problems by proposing a three-branch attention guided convolution neural network (AG-CNN). AG-CNN 1) learns from disease-specific regions to avoid noise and improve alignment, 2) also integrates a global branch to compensate the lost discriminative cues by local branch. Specifically, we first learn a global CNN branch using global images. Then, guided by the attention heat map generated from the global branch, we inference a mask to crop a discriminative region from the global image. The local region is used for training a local CNN branch. Lastly, we concatenate the last pooling layers of both the global and local branches for fine-tuning the fusion branch. The Comprehensive experiment is conducted on the ChestX-ray14 dataset. We first report a strong global baseline producing an average AUC of 0.841 with ResNet-50 as backbone. After combining the local cues with the global information, AG-CNN improves the average AUC to 0.868. While DenseNet-121 is used, the average AUC achieves 0.871, which is a new state of the art in the community.

• The paper introduces AG-CNN, which innovatively merges global and local image features to enhance thorax disease classification.
• It employs an attention mechanism with heat maps to isolate disease-specific regions, achieving an average AUC of 0.871 on the ChestX-ray14 dataset.
• This method improves small lesion detection and promises clinical diagnostic benefits by reducing noise from irrelevant image areas.

Analysis of "Diagnose like a Radiologist: Attention Guided Convolutional Neural Network for Thorax Disease Classification"

The paper, "Diagnose like a Radiologist: Attention Guided Convolutional Neural Network for Thorax Disease Classification," very effectively addresses the limitations of global image-based CNN methodologies previously employed in thorax disease classification using chest X-rays. Traditional techniques often suffer from noise introduced by non-disease specific areas in image data, compounded by alignment issues due to irregular image borders. The authors propose an attention-guided convolutional neural network (AG-CNN) that innovatively combines local lesion-specific information with global image data to enhance classification performance.

The AG-CNN framework comprises three key branches: a global branch that learns from the entire image, a local branch focused on disease-specific regions, and a fusion branch that integrates the outputs of the former two. The attention mechanism within the local branch utilizes a heat map approach to identify and crop relevant regions from the global image, thereby targeting lesion-specific regions necessary for accurate classification.

Their experimental results, conducted on the well-established ChestX-ray14 dataset, reveal significant improvements in classification accuracy. Specifically, AG-CNN achieves an average AUC of 0.871 with DenseNet-121 serving as the backbone, which is notably superior to previous state-of-the-art methods, marking a 2.9% increase over the best results reported prior. This enhancement is substantial given the already competitive baselines like ResNet-50, which posted average AUC results of 0.841.

The paper poignantly highlights that smaller lesions, such as nodules, benefit greatly from the local attention capability, with AG-CNN improving classification beyond what purely global methods could achieve. This result aligns with radiological practice, where nuanced regions, often inconspicuous in global contexts, hold key diagnostic importance.

The introduction of mask inference, which uses activation values to generate attention heat maps, stands as a pivotal element. These maps successfully locate high-saliency regions pertinent for disease identification, adhered by optimizing threshold parameters. Such methodology elegantly merges clinical intuitiveness with technical precision, reminiscent of the radiologist’s approach to focusing on localized pathologies.

Practically, AG-CNN holds promising implications for the integration into clinical diagnostic workstations, potentially acting as a decision support tool to mitigate human error amid the complexity of reading X-rays. Its emphasis on masking irrelevant data while enhancing lesion focus could significantly expedite the radiological diagnostic process, maintaining desired accuracy levels.

Regarding future directions, the authors indicate potential advancements through improved precision in lesion localization and exploring semi-supervised learning paradigms. Such explorations would address the scarcity of annotated medical image data and continue optimizing AG-CNN for broader clinical applicability.

Overall, the paper advances the field of medical image analysis by introducing a refined mechanism to simultaneously leverage local and global image features, addressing existing challenges in thoracic radiology classification tasks. Its contributions provide insights into enhancing diagnostic technology and pave avenues for further research endeavors in medical image processing and deep learning.