PaperlayerView exact revision on arXiv

Source-linked AI summary

Automatic Segmentation and Disease Classification Using Cardiac Cine MR Images

Jelmer M. Wolterink, Tim Leiner, Max A. Viergever, Ivana Isgum

arXiv:1708.01141v1cs.CV

TL;DR

Cardiac cine MR segmentation is challenging, yet it provides quantitative indices used to assess cardiac function and disease. This paper combines CNN segmentation with Random Forest diagnosis and reports 91% multi-class accuracy, while noting that only ED and ES images were used.

  • Problem

    Segmenting cine MR is challenging, while segmentations provide quantitative indices used to quantify cardiac function and identify cardiac disease.

    Source: §1 · p. 1

  • Method

    A CNN segments the LV, RV, and myocardium, and a Random Forest classifies patients using derived features and patient information.

    Source: §1 · p. 2

  • Results

    91% multi-class accuracy was achieved, and the automatically obtained segmentations had good overlap with manual references.

    Source: §5 · pp. 8–9

  • Takeaways & Limitations

    The classifier’s posterior probabilities can identify patients that are difficult to assign to a single disease class.

    Source: §5 · p. 9

    The study used only ED and ES images; whether the full cine sequence could improve segmentation or diagnosis remains for future work.

    Source: §5 · p. 9

Abstract

from arXiv · show

Segmentation of the heart in cardiac cine MR is clinically used to quantify cardiac function. We propose a fully automatic method for segmentation and disease classification using cardiac cine MR images. A convolutional neural network (CNN) was designed to simultaneously segment the left ventricle (LV), right ventricle (RV) and myocardium in end-diastole (ED) and end-systole (ES) images. Features derived from the obtained segmentations were used in a Random Forest classifier to label patients as suffering from dilated cardiomyopathy, hypertrophic cardiomyopathy, heart failure following myocardial infarction, right ventricular abnormality, or no cardiac disease. The method was developed and evaluated using a balanced dataset containing images of 100 patients, which was provided in the MICCAI 2017 automated cardiac diagnosis challenge (ACDC). The segmentation and classification pipeline were evaluated in a four-fold stratified cross-validation. Average Dice scores between reference and automatically obtained segmentations were 0.94, 0.88 and 0.87 for the LV, RV and myocardium. The classifier assigned 91% of patients to the correct disease category. Segmentation and disease classification took 5 s per patient. The results of our study suggest that image-based diagnosis using cine MR cardiac scans can be performed automatically with high accuracy.

1 Introduction

Cardiac cine MR segmentations support quantitative indices used to identify and monitor cardiac disease, but producing accurate segmentations is challenging. The paper proposes automatic segmentation and disease classification from cine MR images.

1 Introduction

  • Quantitative indices such as ejection fraction and myocardial mass are extracted from ventricular and myocardium segmentations to identify cardiac disease.The passage also identifies their use in diagnosis and monitoring.Source: §1 · p. 1
  • Cine MR segmentation is challenging because images are highly anisotropic, may have poor contrast, and disease can cause large anatomical variation.Source: §1 · p. 1
  • Accurate cine MR segmentation remains an ongoing endeavor, with recent contributions from deep learning methods.Source: §1 · p. 2
  • The authors combine fully automatic LV, RV, and myocardium segmentation with quantitative features and patient information in a Random Forest for disease classification.They also describe detecting patients with ambiguous indications.Source: §1 · p. 2

2 Data

The study uses the 100 labeled training cases from the ACDC challenge, which provides cine MR scans across five diagnostic classes. The cases include ED and ES reference segmentations, with image spacing and intensity preprocessing applied.

2 Data

  • ACDC provides 150 patients across five classes, with reference standards for 100 training cases and no reference standards for 50 test cases; this study uses the training cases.The classes are normal, DCM, HCM, MINF, and RVA.Source: §2 · p. 2
  • Each patient has 12–35 short-axis cine MR frames spanning the LV, with indicated ED and ES frames and manual LV, RV, and myocardium segmentations at both phases.In-plane spacing is 1.37–1.68 mm and inter-slice spacing is 5–10 mm.Source: §2 · p. 2
  • Slices were resampled to 1.4 × 1.4 mm2 and each MR volume was intensity-normalized using its 5th and 95th percentiles.Intensities were normalized to [0.0, 1.0].Source: §2 · p. 2

3 Methods

The pipeline uses a CNN to segment LV, RV, and myocardium in short-axis cine MR slices, then classifies patients with a Random Forest using segmentation-derived indices and patient information. Segmentation uses paired ED/ES inputs, jointly optimized outputs, and snapshot ensembling.

3 Methods

  • A CNN segments the LV cavity, RV cavity, and myocardium, and a Random Forest combines indices from the segmentations with patient information to assign one of five diagnoses.Source: §3 · p. 2

3.1 Segmentation

  • The segmentation CNN uses dilated convolutional layers to obtain a 131 × 131 voxel receptive field while retaining few parameters and high-resolution feature maps.Batch Normalization was used to mitigate potential overfitting.Source: §3.1 · p. 3
  • Anatomically aligned ED and ES slices are input together, and the CNN predicts separate four-class probability maps for each phase before assigning voxel labels.The four labels are background, LV, RV, and myocardium.Source: §3.1 · p. 3, Figure 1 · p. 3
  • Training minimizes soft Dice loss across all eight phase-specific classes, with the coefficient based on reference masks and predicted probability maps.The class-wise coefficients are averaged to jointly optimize all classes.Source: §3.1 · pp. 3–4
  • Snapshot ensembling stores model copies at cyclic learning-rate resets; testing averages six stored models’ probability maps before assigning labels.Training ran for 150,000 iterations, with resets every 10,000 iterations; only the largest 3D 6-connected component was retained per class at testing.Source: §3.1 · p. 4

3.2 Diagnosis

  • The diagnosis Random Forest uses 14 features—two patient characteristics and 12 image-derived indices—and comprises 1,000 full-depth trees.Posterior probabilities determine the assigned class, and their entropy estimates classifier uncertainty.Source: §3.2 · p. 4

4 Experiments and Results

The four-fold stratified evaluation reports 91% diagnosis accuracy and segmentation agreement that varies by chamber and cardiac phase. It also identifies influential classification features and cases with high classification uncertainty.

4 Experiments and Results

  • Evaluation used stratified four-fold cross-validation on 100 training cases, with 15 patients per class for training and five per class for testing in each fold.Source: §4 · pp. 4–5

4.1 Segmentation Results

4.2 Diagnosis Results

  • 91% overall accuracy was achieved across five diagnosis classes, with sensitivity of 100% for normal, 90% for DCM/HCM/RVA, and 85% for MINF.Source: Table 2 · p. 7, §4.2 · p. 7

4.1 Segmentation Results

  • LV EF correlated with reference values at 0.97 versus 0.86 for RV EF, alongside slight LV/RV underestimation and myocardium overestimation at both phases.Source: §4.1 · pp. 5, 7, Figure 3 · p. 6

4.2 Diagnosis Results

  • CNN segmentation took 4 seconds per patient, and feature extraction plus Random Forest classification took around 1 second.Source: §4.1 · p. 7, §4.2 · p. 8
  • The three most important classifier features were LV EF, RV-to-LV volume ratio at ED, and myocardium-to-LV volume ratio at ES.Source: §4.2 · p. 7, Figure 4 · p. 8
  • Several patients remained uncertain: Patient 18’s DCM was classified as MINF, while Patient 44’s MINF was classified as normal.Source: §4.2 · pp. 7–8, Figure 5 · p. 8

5 Discussion and Conclusion

The method achieves good overlap with manual segmentations and 91% multi-class diagnostic accuracy. The authors discuss how the classifier relates to clinical guidelines, potential effects of segmentation errors, and the study’s use of only ED and ES images.

5 Discussion and Conclusion

  • 91% multi-class accuracy was achieved using the automatically obtained segmentations.The segmentations of the left ventricle, right ventricle, and myocardium also showed good overlap with manual references.Source: §5 · pp. 8–9
  • The Random Forest partially captures clinical guidelines, and its posterior probabilities can flag patients difficult to assign to one disease.Whether classifier uncertainty corresponds to uncertainty in clinical diagnosis remains for future investigation.Source: §5 · p. 9
  • Segmentation errors could affect classification, but reference-based indices improved accuracy only from 91% to 92% at considerable added time and effort.The authors also report very small bias in quantitative indices derived from reference versus automatic segmentations.Source: §5 · p. 9
  • Only ED and ES images were used; whether complete-sequence inputs improve segmentation or diagnosis remains untested.Source: §5 · p. 9

Related papers

Loading 1708.01141v1…