Fully Convolutional Multi-scale Residual DenseNets for Cardiac Segmentation and Automated Cardiac Diagnosis using Ensemble of Classifiers (1801.05173v1)

Abstract: Deep fully convolutional neural network (FCN) based architectures have shown great potential in medical image segmentation. However, such architectures usually have millions of parameters and inadequate number of training samples leading to over-fitting and poor generalization. In this paper, we present a novel highly parameter and memory efficient FCN based architecture for medical image analysis. We propose a novel up-sampling path which incorporates long skip and short-cut connections to overcome the feature map explosion in FCN like architectures. In order to processes the input images at multiple scales and view points simultaneously, we propose to incorporate Inception module's parallel structures. We also propose a novel dual loss function whose weighting scheme allows to combine advantages of cross-entropy and dice loss. We have validated our proposed network architecture on two publicly available datasets, namely: (i) Automated Cardiac Disease Diagnosis Challenge (ACDC-2017), (ii) Left Ventricular Segmentation Challenge (LV-2011). Our approach in ACDC-2017 challenge stands second place for segmentation and first place in automated cardiac disease diagnosis tasks with an accuracy of 100%. In the LV-2011 challenge our approach attained 0.74 Jaccard index, which is so far the highest published result in fully automated algorithms. From the segmentation we extracted clinically relevant cardiac parameters and hand-crafted features which reflected the clinical diagnostic analysis to train an ensemble system for cardiac disease classification. Our approach combined both cardiac segmentation and disease diagnosis into a fully automated framework which is computational efficient and hence has the potential to be incorporated in computer-aided diagnosis (CAD) tools for clinical application.

• The paper presents a novel fully convolutional multi-scale residual DenseNet that efficiently addresses cardiac MRI segmentation challenges.
• It introduces a dual loss function combining cross-entropy and dice loss, achieving a Jaccard index of 0.74 on the LV-2011 dataset.
• The integrated framework for segmentation and automated diagnosis demonstrated 100% accuracy on the ACDC dataset, offering clinical viability.

Overview of Fully Convolutional Multi-scale Residual Dense Networks for Cardiac Segmentation and Automated Diagnosis

The paper presents a novel framework designed for cardiac segmentation and automated cardiac diagnosis leveraging advanced deep learning techniques. The proposed architecture, a fully convolutional multi-scale residual DenseNet, is highly parameter and memory efficient. This framework addresses a significant challenge in medical image segmentation, particularly for cardiac MRI, where deep network architectures traditionally face difficulties due to limited training samples, overfitting, and computational overhead.

Key Contributions

1. Efficient Network Architecture: The researchers designed a fully convolutional network that incorporates both long skip and short-cut connections. This design mitigates the common issue of feature map explosion associated with FCN architectures. The network processes input images at multiple scales using parallel structures inspired by the Inception module. This capability enables the network to efficiently capture and process varying scales and viewpoints inherent in medical images.
2. Dual Loss Function: A novel dual loss function combines the benefits of cross-entropy and dice loss, each weighted to optimize both pixel-wise accuracy and segmentation metrics. The careful weighting scheme demonstrates adaptability to different classes within the dataset, addressing problems of class imbalance typical in medical imaging tasks.
3. Comprehensive Validation: Validation of the proposed architecture on the ACDC-2017 and LV-2011 datasets yielded competitive performance. Specifically, the architecture achieved a high Jaccard index of 0.74 on the LV-2011 dataset, the highest among fully automated methods, while also surpassing other methods in automatic cardiac disease diagnosis with an accuracy rate of 100% on the ACDC testing dataset.
4. Integrated Framework for Diagnosis: Beyond segmentation, this framework also integrates cardiac structure segmentation with the classification of cardiac diseases into a holistic automated solution. This approach is computationally efficient, lending itself to clinical applications where there is a demand for rapid and reliable computer-aided diagnosis tools.

Implications and Future Directions

Practically, this research demonstrates a clear pathway for deploying deep learning models in clinical environments, especially given the network's reduced parameter requirements and computational efficiency. The use of high-dimensional residual DenseNet connections enhances model robustness to overfitting, a critical consideration given the typical scarcity of annotated medical training data.

Theoretically, the paper contributes to the growing body of knowledge on how modifications in network architecture—such as employing inception-style parallel convolution branches—can enhance multi-scale image analysis capabilities within deep learning models. There is potential for further exploration into extending this dense connectivity pattern and multi-scale processing paradigm to 3D volumes, offering advances in volumetric segmentation tasks across other domains of medical imaging.

In summary, while further clinical validation and real-world testing are essential, the paper lays essential groundwork for transforming cardiac diagnostics via deep learning, with implications extending beyond just cardiac imaging. Future work can expand on this foundation, exploring more complex architectures and real-time deployment in busy clinical settings to further advance computer-aided diagnosis.