Fully Convolutional Architectures for Multi-Class Segmentation in Chest Radiographs (1701.08816v4)

Abstract: The success of deep convolutional neural networks on image classification and recognition tasks has led to new applications in very diversified contexts, including the field of medical imaging. In this paper we investigate and propose neural network architectures for automated multi-class segmentation of anatomical organs in chest radiographs, namely for lungs, clavicles and heart. We address several open challenges including model overfitting, reducing number of parameters and handling of severely imbalanced data in CXR by fusing recent concepts in convolutional networks and adapting them to the segmentation problem task in CXR. We demonstrate that our architecture combining delayed subsampling, exponential linear units, highly restrictive regularization and a large number of high resolution low level abstract features outperforms state-of-the-art methods on all considered organs, as well as the human observer on lungs and heart. The models use a multi-class configuration with three target classes and are trained and tested on the publicly available JSRT database, consisting of 247 X-ray images the ground-truth masks for which are available in the SCR database. Our best performing model, trained with the loss function based on the Dice coefficient, reached mean Jaccard overlap scores of 95.0\% for lungs, 86.8\% for clavicles and 88.2\% for heart. This architecture outperformed the human observer results for lungs and heart.

• The paper introduces three novel CNN architectures—All-Dropout, All-Convolutional, and InvertedNet—that significantly enhance multi-class segmentation in chest radiographs.
• The InvertedNet model achieved Jaccard scores of 95.0% for lungs, 86.8% for clavicles, and 88.2% for heart, surpassing performance of human observers.
• Methodological innovations such as replacing max-pooling with strided convolutions and strategic dropout address challenges in boundary detection and class imbalance.

Fully Convolutional Architectures for Multi-Class Segmentation in Chest Radiographs

The automated segmentation of anatomical structures in chest radiographs (CXR) is a well-researched area of medical imaging, offering significant potential for aiding radiological diagnostics. This paper introduces comprehensive convolutional neural network (CNN) architectures optimized for multi-class segmentation of lungs, clavicles, and heart, challenging conventional methodologies by emphasizing fully convolutional network designs. Traditional CXR segmentation methods often face limitations due to anatomical variations that impact boundary detection.

Methodology

The paper describes three main architectures that aim to enhance segmentation accuracy in CXR: All-Dropout, All-Convolutional (All-Conv), and InvertedNet. Each adopts a unique approach to overcome challenges such as overfitting and class imbalance:

1. All-Dropout: This architecture intensively utilizes dropout layers after each convolutional layer. By Gaussian noise addition or neuron dropout, it strengthens regularization to prevent overfitting. The dropout layers regularly decrease the parameter space's freedom, thereby establishing a more robust generalization on unseen data.
2. All-Convolutional: Following the approach of Springenberg et al., this variant replaces the conventional max-pooling layers with convolutional layers that operate at an increased stride. This replacement encourages pooled spatial data without unnecessary loss of precision, thereby maintaining a higher resolution of learned features throughout the network.
3. InvertedNet: Optimizing for complexity reduction, InvertedNet employs delayed subsampling and reduces the feature map size progressively while maintaining a high resolution of low-level abstract features. This inversion of feature map considerations leads to fewer parameters overall but maximizes detailed feature learning essential for precise boundary detection in diverse anatomical structures.

Training across these architectures incorporates strategies such as using loss functions weighted by class distribution to counteract the data imbalance inherent in the segmentation task since lung areas dominate pixel-wise in CXR images.

Results and Performance

The InvertedNet model showed particularly promising results, achieving Jaccard overlap scores of 95.0% for lungs, 86.8% for clavicles, and 88.2% for heart segmentation on the JSRT dataset. These results demonstrate a substantial improvement over both state-of-the-art methodologies and human observers for lungs and heart segmentation. Human observer results noted overlap scores of 94.6% for lungs and 87.8% for the heart, while the algorithm exceeded these values, illustrating its effectiveness and robustness.

Implications and Future Work

The implications of these findings are prominent in advancing automated diagnostic tools. Especially, these networks improve the accuracy and efficiency needed in high-volume clinical settings where hundreds of CXRs may need processing daily. The reduction in computational cost and increased efficiency brought by such streamlined architectures demonstrate compelling advantages for real-world adoption.

Future investigations highlighted in this paper will focus on further refining these architectures through the integration of more complex networks such as those including DenseNet-like connections to facilitate enhanced gradient flow. Additionally, transferring knowledge from linked tasks via deep/transfer learning also holds potential to mitigate the data scarcity challenge in medical imaging.

Conclusion

The innovations introduced within this paper provide meaningful advancements in the application of CNNs for CXR segmentation tasks. While the lung field results have approached saturation, clavicle segmentation remains challenging, signifying an area for ongoing development. By balancing architectural complexity and model accuracy, this research shines light on practical, high-precision automated tools with substantial clinical value. Moving forward, the paper paves the path to explore architectural enhancements and integration with broader imaging diagnostics, indicating a fruitful direction for future work in computer-aided diagnostics.