Abstract
Pneumonia remains a significant global health threat, necessitating efficient and accurate diagnostic tools. Automated detection using Chest X-ray (CXR) images via deep learning offers a promising path to reduce clinical workloads. However, standard object detection architectures often struggle with the balance between high-level semantic extraction and the preservation of spatial resolution. In this paper, we present GeminiNet, a specialized framework that integrates a Fully Convolutional Network (FCN) with a novel Receptive Field Block (RFB). By utilizing a DetNet59 backbone and a retina-like convolutional structure, GeminiNet expands the receptive field to capture multi-scale features without losing critical spatial information. Evaluated on the RSNA dataset, GeminiNet achieved an Average Precision (AP) of 0.4575 and an inference speed of 8 fps, outperforming standard ResNet-based Faster R-CNN models.
Introduction
The diagnosis of pneumonia from Chest X-rays (CXR) is a high-stakes task where accuracy and speed are paramount. While deep learning has revolutionized medical imaging, standard architectures often face a trade-off: deep layers capture global context but lose fine-grained local details due to downsampling, while shallow layers lack the semantic depth needed for complex classification. GeminiNet addresses this by rethinking the receptive field—the specific area of an input image that influences a given feature.
The Resolution-Localization Trade-off
Current state-of-the-art models for object detection face two primary technical bottlenecks in medical imaging:
- Resolution Loss: Backbones like ResNet utilize heavy downsampling. In deep layers, the semantic features of small targets, such as early-stage pneumonia opacities, can effectively disappear.
- The Fully Connected Bottleneck: Many detectors rely on fully connected (FC) layers for final classification. FC layers are computationally expensive and can discard spatial information necessary for precise bounding box localization.
The GeminiNet Architecture
GeminiNet is built upon the concept of Region-based Fully Convolutional Networks (R-FCN) but introduces several structural innovations to overcome the aforementioned bottlenecks.
DetNet59 Backbone
Instead of a standard classification backbone, GeminiNet employs DetNet59. This backbone is specifically designed for object detection, maintaining a higher feature map resolution in its later stages compared to traditional networks.
Receptive Field Block (RFB)
The core innovation of GeminiNet is the Receptive Field Block (RFB), integrated into the sixth stage of the network. Inspired by the human visual system, the RFB uses an Inception-like design with parallel branches of different kernel scales ($1 \times 1$, $1 \times 3$, $3 \times 1$, and $3 \times 3$) combined with dilated convolutions.
Dilated convolutions allow the network to "skip" pixels, effectively expanding the field of view to capture global context without increasing the parameter count or reducing resolution. This "retina-like" structure enables the model to process information at multiple scales simultaneously.
Dual-Branch Information Flow
The "Gemini" designation refers to the dual-stream processing of information. The network utilizes Position Sensitive Region of Interest (PSRoI) Pooling to capture local, part-based information and RoI Align to capture global context. By combining these streams, the model preserves the spatial "vote" of different image parts, ensuring translation variance—a critical requirement for accurate object localization.
Experimental Setup and Data
The model was rigorously tested using the RSNA Pneumonia Detection Challenge dataset, a subset of the NIH CXR14 dataset containing 30,000 exams.
To address data scarcity and improve robustness, the authors implemented data augmentation through horizontal and vertical flipping. This expanded the training set and helped the model generalize better to variations in patient positioning.
Performance Analysis
GeminiNet demonstrated superior performance across several key metrics on the RSNA dataset:
- Accuracy: At an IoU (Intersection over Union) threshold of 0.5, GeminiNet reached an Average Precision (AP) of 0.4575, an improvement of 0.078 over a standard ResNet50 baseline.
- Classification Power: The model achieved an AUC of 0.7758, indicating high reliability in distinguishing between positive and negative pneumonia cases.
- Efficiency: GeminiNet achieved a detection speed of 8 fps, outperforming the 7 fps typical of the popular Faster R-CNN architecture.
- Comparative Gains: The architecture showed a 10.7% improvement in Mean Average Precision (mAP) over Mask R-CNN.
Conclusion
GeminiNet represents a significant step forward in automated pneumonia detection. By combining a detection-optimized backbone with a multi-scale receptive field block, it successfully balances the need for high-resolution spatial data with deep semantic context. While future work will explore attention mechanisms to further refine the weighting between local and global features, GeminiNet currently offers a high-performance, efficient solution for assisted medical reading.
Summary updated