PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Mahendra Khened, Varghese Alex Kollerathu, Ganapathy Krishnamurthi

arXiv:1801.05173v1cs.CV

TL;DR

Manual cardiac analysis is time-consuming and variable, while fully convolutional architectures can be difficult to train under class imbalance and may use many parameters. The paper presents a parameter-efficient FCN with residual skip connections and reports near-state-of-the-art segmentation results, including 100% ACDC diagnostic accuracy and 0.4 million parameters. Its main limitation is difficulty segmenting apex and basal cardiac slices.

  • Problem

    Manual analysis is time-consuming and introduces intra- and inter-rater variability, while class imbalance makes training difficult.

  • Method

    The paper proposes a novel 2D FCN architecture using residual long skip and shortcut connections in the up-sampling path.

  • Results

    The approach achieved 100% accuracy on the ACDC testing dataset, near-state-of-the-art segmentation results, and used 0.4 million parameters.

  • Takeaways & Limitations

    The reported results support a computationally efficient approach for cardiac segmentation and diagnosis.

  • Takeaways & Limitations

    The model cannot segment cardiac structures reliably in extremely difficult apex and basal heart slices.

Abstract

from arXiv · show

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.

1. Introduction

The paper presents a parameter- and memory-efficient 2D FCN for cardiac image segmentation and a fully automated framework linking segmentation with cardiac disease diagnosis. Its design combines multi-scale processing, efficient up-sampling connections, a dual loss, and ensemble classification, with validation on two cardiac challenges.

  • Manual cardiac contouring is time-consuming, tedious, and subject to intra- and inter-rater variability, motivating fully automatic segmentation and diagnosis.
  • The proposed 2D FCN is designed to be highly parameter- and memory-efficient while supporting automated cardiac structure segmentation and disease diagnosis.
  • Parallel convolutions with different kernel sizes provide multi-scale processing, while long skip and short-cut connections make the up-sampling path more computationally and memory efficient.
  • A weighted combination of cross-entropy and Dice loss is introduced to target pixel-wise accuracy and segmentation performance.
  • Segmentation labels support clinical and hand-crafted feature extraction, followed by Random Forest feature-importance analysis and a two-stage ensemble classifier for cardiac disease prediction.

2. Material and Methods

The paper presents a fully automated cardiac MRI pipeline combining ROI extraction, multi-scale dense segmentation, and ensemble-based disease diagnosis. Experiments use the ACDC-2017 and LV-2011 datasets, while the proposed architecture targets parameter, memory, and feature-map efficiency.

  • Loss and diagnosis: Disease diagnosis uses an ensemble of classifiers trained on clinically relevant parameters and hand-crafted features extracted from predicted segmentation labels.The approach includes a two-class expert classifier trained only on myocardial-wall information.
  • Pipeline and datasets: The pipeline crops a 128 × 128 ROI around the approximate LV center using Fourier analysis and the Circular Hough Transform before CNN training and inference.ROI extraction helps alleviate class imbalance associated with heart-structure labels in full-sized cardiac MR images.
  • Segmentation architecture: Replacing skip-connection concatenation with projection and element-wise addition reduces parameters and GPU memory footprint without affecting segmentation quality.The architecture also adds residual short-cut connections in the up-sampling path and parallel pathways in the initial layer.
  • Segmentation architecture: The final DFCN-C architecture incorporates the modifications introduced in DFCN-B and adds parallel pathways to its initial layer.The proposed modifications were intended to improve parameter efficiency, convergence rate, and memory footprint.
  • Loss and diagnosis: A dual loss function combines cross-entropy and Dice loss, with additional weighting mechanisms for training the segmentation network.The network parameters are optimized against both losses in tandem, with an L2 weight-decay penalty.

3. Experimental analysis

The experiments evaluate the proposed architecture, loss, preprocessing, augmentation, and feature-based classification components. Results indicate that multi-scale feature extraction, feature fusion, and efficient connectivity support segmentation performance with few parameters.

  • Experimental setup: The experiments assessed architecture, loss function, ROI cropping, post-processing, data augmentation, and classification features using Dice score and Hausdorff Distance.Ablation studies used the ACDC training dataset, while the network was implemented in TensorFlow.
  • Network configuration: The network used three max-pooling operations, growth-rate k = 12, and at most approximately three times k initial feature maps.These hyper-parameters were fixed based on experimental results.
  • Growth-rate analysis: Increasing the growth-rate steadily improved segmentation performance while retaining extremely few trainable parameters.This relationship was evaluated for the same DFCN-C architecture.
  • Preprocessing and post-processing: ROI cropping and post-processing were selected to reduce computational demands and remove false positives, despite non-ROI images achieving a better Dice score before these considerations.Non-ROI processing had higher Hausdorff distance because of false positives at basal and apical slices.
  • Augmentation and architecture comparison: Data augmentation produced a minor validation improvement corroborated on the held-out set, while DFCN-C showed faster convergence and better scores than DFCN-B.The comparison supports the proposed multi-scale feature extraction and feature-fusion design.

4. Results

The proposed framework was evaluated for cardiac segmentation and automated disease diagnosis across ACDC-2017 and LV-2011 settings. It achieved strong comparative performance and generalized across datasets, while errors remained concentrated in challenging slices and anatomies.

  • Evaluation measures: The evaluation measured clinical volume, ejection-fraction, and myocardium-mass errors, along with Dice, Hausdorff, and disease-diagnosis accuracy.Segmentation metrics covered LV, RV, and MYO at ED and ES phases.
  • ACDC segmentation: The method ranked 2nd for LV and MYO segmentation and 3rd for RV in the ACDC challenge.The reported ranking was based on the challenge leaderboard.
  • ACDC segmentation: Basal and apical slices sometimes produced erroneous segmentations, including basal-slice RV errors and apical-slice myocardium over-segmentation.Most slices were segmented accurately, but these errors limited performance in difficult regions.
  • Automated diagnosis: The two-stage diagnosis approach with hand-crafted myocardial wall-thickness features surpassed other methods in automated cardiac disease diagnosis.Compared methods mostly relied on cardiac volumetric features and single-stage classification.
  • LV-2011 segmentation: 0.68 ± 0.16, 0.78 ± 0.13, and 0.74 ± 0.18 were the Jaccard indices for apex, mid, and base slices, respectively, on LV-2011.Errors were mostly concentrated in apical slices, while the approach was on par with other fully automated methods across most measures.
  • Cross-dataset generalization: Training on ACDC enabled segmentation of all cardiac frames in LV-2011 despite using only ED and ES frames and one-sixth of the LV-2011 training images.The result was presented as evidence of generalization across datasets with different data distributions.

5. Discussion and conclusion

The paper presents a compact cardiac MR segmentation and diagnosis framework designed for efficient processing and improved generalization on small datasets. Its network combines residual connectivity, Inception-style parallelism, and dual-loss training, achieving near state-of-the-art segmentation while supporting automated diagnosis.

  • 5. Discussion and conclusion: The 2D-DFCN uses Residual DenseNets with Inception architecture for cardiac MR segmentation.The design targets efficient processing while learning cardiac structures across short-axis slices and cardiac phases.
  • 5. Discussion and conclusion: The framework achieved near state-of-the-art segmentation on two benchmark cardiac MR datasets and combined segmentation with automated disease diagnosis.The pipeline was evaluated across clinical-institution variability and included ROI cropping, CNN segmentation, and downstream analysis.
  • 5. Discussion and conclusion: 0.4 million trainable parameters represented an order of 10-fold reduction compared with standard U-Net-based architectures.The authors report this as the lowest parameter count among most compared CNN-based cardiac MR approaches.
  • 5. Discussion and conclusion: The network showed better generalization on small datasets without data augmentation, whereas FCN and U-Net architectures tended to over-fit.The authors attribute this performance to the proposed connectivity pattern and low model capacity.
  • 5. Discussion and conclusion: Residual long skip and short-cut connections reduce memory explosion in the up-sampling path.These connections address the feature-map growth associated with DenseNet-based FCN designs.
  • 5. Discussion and conclusion: The model has potential for volumetric segmentation and clinical computer-aided diagnosis, while segmentation remains difficult in apical and basal slices.Apical errors were reported to have minor impact on overall volume computation for cardiac disease diagnosis.

Appendix A. Evaluation Metrics

The appendix defines segmentation metrics by comparing predicted and ground-truth volumes or contours. It covers overlap, classification, and contour-distance measures, including their ranges and spatial units.

  • Dice overlap coefficient measures similarity between predicted segmentation maps and ground truth, ranging from zero to one for perfect overlap.The binary-volume form uses predicted and ground-truth voxel sets and their cardinalities.
  • Jaccard index, also called Intersection over Union, measures similarity and diversity between two sets and varies from zero to one.
  • Sensitivity, specificity, positive predictive value, and negative predictive value are used as classification-oriented evaluation measures.
  • Hausdorff distance measures the symmetric distance between two contours; higher values indicate poorer contour matching.It is computed in millimeters using spatial resolution from the DICOM Pixel Spacing tag.

Appendix B.1. Fourier Analysis

The Fourier-analysis procedure extracts cardiac-motion information from short-axis MR image sequences. It isolates the first harmonic, reconstructs a spatial magnitude volume, and suppresses low-intensity noise.

  • The N-D Fourier transform is equivalent to applying one-dimensional transforms along each array dimension.The transform is defined for an N-D array using roots of unity.
  • Short-axis cardiac MR images across the cardiac cycle form a 2D+T signal with Height × Width × Time dimensions.Myocardium and ventricles change significantly over time because of heart motion.
  • The first harmonic, or H1 component, identifies pixel regions with the strongest response to cardiac frequency.These regions correspond to structures responsible for the heartbeat.
  • The H1 component is inverse-transformed into the spatial domain, and its magnitude is retained while phase is discarded.H1 components from base-to-apex slices are stacked into a 3-D volume.
  • Pixel values below 1% of the whole-volume maximum intensity are discarded to reduce noise.

Appendix B.2. Circular Hough Transform

The method localizes the left ventricle by detecting circular myocardial boundaries in H1 component images. It uses edge detection, Gaussian-kernel Hough voting, and likelihood-based center selection.

  • H1 component images show cardiac-motion regions as bright areas, and Canny edges reveal two concentric circles approximating myocardial boundaries at End Diastole and End Systole.
  • The resulting localization uses a Gaussian Kernel-based Circular Hough Transform approach.
  • For each slice, circular Hough transforms search a range of radii and retain the P highest-scoring circles.
  • Gaussian-kernel voting models uncertainty in each circle center and makes the transform more robust to spurious circles.
  • The maximum of the left-ventricle likelihood surface defines the ROI center, from which a fixed-size 128 × 128 square patch is cropped.

Appendix C.1. Overview of DenseNets

DenseNets reuse feature maps through dense-block connectivity, while the described residual connections add identity shortcuts for deep-network training. The architecture emphasizes feature reuse, efficiency, and direct supervision.

  • DenseNet connectivity: Dense blocks iteratively concatenate preceding feature maps, passing them to subsequent layers.Each layer receives preceding feature maps with matching spatial resolution.
  • Efficiency: A small growth rate, typically k=8, makes the number of feature maps grow linearly with network depth.
  • Layer design: Each DenseNet layer uses BN, ELU, convolution, and dropout, while Transition Down applies 1 × 1 convolution followed by 2 × 2 max-pooling.
  • Efficiency: DenseNet connectivity provides higher parameter and computation efficiency than normal ConvNets by using fewer channels per layer.The passage contrasts O(C × C) parameter scaling in normal ConvNets with O(lth × k × k) in DenseNets.
  • Feature reuse: Feature reuse consolidates information from preceding layers, producing more diversified and pattern-rich features.
  • Limited-data setting: Dense connectivity can preserve low- and high-complexity features when training data is minimal and supports smoother decision boundaries.
  • Residual connections: Residual blocks use identity mappings and element-wise summation, reformulating the layer output as H(x) = F(x) + x.The residual function F(x) is learned instead of directly approximating the full transformation H(x).

Appendix C.3. Overview of Inception Architectures

The paper adapts Inception-style parallel processing for semantic segmentation, combining multiple receptive-field scales while limiting computational and feature-map costs.

  • The original architecture stacks Inception modules with occasional stride-2 max-pooling and uses 1 × 1 convolutions for dimension reduction.The original modules were used only in higher layers, while lower layers retained traditional convolutions.
  • Inception modules use parallel paths with different filter sizes to process visual information at multiple scales simultaneously.These paths are aggregated so later stages can abstract features from different scales.
  • The ratio of 3 × 3, 5 × 5, and 7 × 7 convolutions can be skewed, such as 2 : 1 : 1, because larger kernels cover more spatial context.

Appendix D.1. Myocardial Wall Thickness Variation Profile Features

The paper estimates myocardial wall-thickness variation from segmented cardiac contours across short-axis slices and cardiac phases, then uses derived features for disease classification. The analysis exposes difficulty separating DCM from MINF, with MLP classification reaching 100% accuracy on the held-out task.

  • Wall-thickness estimation: Myocardial wall thickness is estimated as the shortest Euclidean distance from an interior contour pixel to any exterior contour pixel.The procedure begins with myocardial segmentation and contour extraction.
  • Wall-thickness estimation: Binary hole-filling, image subtraction, and Canny edge detection produce the cavity and interior and exterior contours used for thickness estimation.Canny edge detection uses sigma = 1.
  • Feature construction: Mean and standard deviation of myocardial wall thickness are estimated for every short-axis slice during end-diastole and end-systole.
  • Feature construction: Eight features quantify maximum, variability, and smoothness of myocardial wall-thickness profiles across short-axis slices and the long axis.Four features are defined for end systole, with a similar set of four for end diastole.
  • Clinical interpretation: Clinically, wall thickness and its variation distinguish MINF from DCM: DCM produces an extremely thin wall, while MINF produces non-uniform thickness.Normal cases show uniform thickness across short-axis slices and smooth long-axis variation.
  • Classification results: 100% MLP accuracy was achieved on the held-out DCM-versus-MINF classification task.Random Forest feature importance favored volumetric features such as EF over myocardial wall-thickness variation features, while classifiers often confused DCM and MINF.
Loading 1801.05173v1…