Source-linked AI summary
2D-3D Fully Convolutional Neural Networks for Cardiac MR Segmentation
Jay Patravali, Shubham Jain, Sasank Chilamkurthy
TL;DR
Manual cardiac MR delineation is time-consuming and error-prone, while slice-wise 2D models can miss spatial context needed to segment the whole heart. The paper develops fully automated 2D and 3D CNN pipelines for cardiac structures, reporting near-state-of-the-art geometric accuracy and convincing clinical performance.
Problem
Manual cardiac delineation is time-consuming and prone to error and rater variability, while slice-wise 2D models miss spatial context needed to segment the whole heart.
Source: §1 · pp. 1–2Method
The paper develops fully automated 2D slice-wise and 3D volumetric CNN models to segment the left ventricle, right ventricle, and myocardium, comparing dice, cross-entropy, and combined losses.
Source: §1 · p. 2Results
Takeaways & Limitations
The authors conclude that their lightweight, modular, runtime-efficient models show viability for practical real-world applications.
Source: §5 · p. 8
Abstract
from arXiv · showhide
In this paper, we develop a 2D and 3D segmentation pipelines for fully automated cardiac MR image segmentation using Deep Convolutional Neural Networks (CNN). Our models are trained end-to-end from scratch using the ACD Challenge 2017 dataset comprising of 100 studies, each containing Cardiac MR images in End Diastole and End Systole phase. We show that both our segmentation models achieve near state-of-the-art performance scores in terms of distance metrics and have convincing accuracy in terms of clinical parameters. A comparative analysis is provided by introducing a novel dice loss function and its combination with cross entropy loss. By exploring different network structures and comprehensive experiments, we discuss several key insights to obtain optimal model performance, which also is central to the theme of this challenge.
1 Introduction
Cardiac MR segmentation can support diagnostic measures, but manual delineation is time-consuming and error-prone. This work develops automated 2D and 3D CNN models for cardiac structures and reports near-state-of-the-art accuracy.
1 Introduction
- Manual delineation is time-consuming and prone to error and rater variability, motivating accurate, reproducible, fully automated cardiac segmentation.Segmenting the ventricles and myocardium can provide diagnostic information such as ejection fraction and myocardial mass.Source: §1 · p. 1
- U-Net builds on fully convolutional networks, using skip connections between contracting and expanding paths to support segmentation with few training samples.Source: §1 · p. 1
- Slice-by-slice 2D models miss spatial context needed to segment the whole heart, and few 3D cardiac CNN methods had achieved satisfactory performance.Source: §1 · p. 2
- Both automated CNN models segment the left ventricle, right ventricle, and myocardium, with near-state-of-the-art geometric accuracy and clinical validity.The 2D model operates slice-by-slice, while the 3D model computes volumetric segmentation.Source: §1 · p. 2
2 Method
The method pairs a U-Net-derived 2D slice model with a volumetric 3D extension, trained on ACDC cardiac MRI studies. It compares weighted cross-entropy, logarithmic Dice loss, and their weighted combination.
2.1 Network Architecture
- The 2D model adapts U-Net with contracting and expanding paths, configurable slice inputs, and a final mapping from 32 feature channels to four classes.Its contracting and expanding paths use convolution, batch normalization, and ReLU blocks, with pooling and up-convolution for downsampling and upsampling.Source: §2.1 · pp. 2–3
- The 3D model replaces 2D operations with 3D counterparts, limits feature maps to 256, and pools only across X and Y to support varying slice counts.Dropout probabilities of 0.5 and 0.3 are used in the last and second-last contracting layers, respectively.Source: §2.1 · p. 3
2.2 Dataset, Preprocessing and Augmentation
- The dataset contains 150 annotated exams: 100 for training and 50 for testing, with short-axis scans at end diastole and end systole.Source: §2.2 · p. 3
- Preprocessing applies CLAHE, percentile-based intensity normalization and clipping, resampling to 1.5x1.5x10 mm spacing, and resizing or cropping to 256x256.Training volumes are resized and cropped to 256x256x12; on-the-fly augmentation rotates images from -15 to +15 degrees and scales them from 0.9 to 1.1.Source: §2.2 · pp. 3–4
2.3 Training
- The 2D model trains slice by slice and the 3D model on whole volumes, using five-fold cross-validation, 300 epochs, and batches of 8 and 4, respectively.Training uses stochastic gradient descent with momentum 0.99 and decays the learning rate by a factor of 10 every 30 epochs.Source: §2.3 · p. 4
2.4 Optimization Function
- Both models apply a pixel-wise softmax to predict class probabilities at each location, then compare predictions with target labels using three optimization functions.Source: §2.4 · p. 4, Cross Entropy Loss · p. 4, Dice Loss · p. 4
Cross Entropy Loss
- Weighted cross-entropy sums pixel-level prediction errors and applies class weights to offset differences in pixel frequency.Source: Cross Entropy Loss · p. 4
Dice Loss
- The proposed Dice loss is a weighted sum of class-specific losses based on a logarithmic Dice score, intended to be easier to optimize and useful when background pixels dominate.Source: Dice Loss · pp. 4–5
3 Results
Across 100 ACDC 2017 training studies, the models were evaluated on distance and clinical metrics; Dice loss led across distance metrics, while the 3D model performed better on clinical parameters.
3 Results
- 3D Hausdorff distances were much higher than 2D distances, which the authors attributed to distant false-positive speckles.Source: §3 · p. 5
- Both models trained in under three hours and generated a whole-phase output in 0.3 seconds, using 1.2 GB GPU memory for 2D and 2 GB for 3D.Training used 4 GB GPU memory for each model.Source: §3 · p. 5
- Dice loss outperformed cross-entropy and combined loss across 2D and 3D distance metrics, with scores near those of [14].The distance metrics were Dice Score and Hausdorff Distance for LV, RV and MYO in ED and ES.Source: §3 · p. 5, Table 1 · p. 6, Table 2 · p. 6
- The 3D model outperformed the 2D model on clinical parameters, although their distance metric scores were fairly similar.For MYO, Dice loss significantly outperformed cross-entropy; LV and RV results were fairly similar between those losses.Source: §3 · p. 5, Table 3 · p. 6, Table 4 · p. 6
- Figures 3 and 4 show model predictions alongside inputs and ground truth, including full-phase slices 0–8.Source: Figure 3 · p. 6, Figure 4 · p. 7
4 Discussion
The discussion examines slice context, augmentation choices and loss functions. It reports that three-slice inputs performed best, lighter augmentation improved scores, and Dice loss improved both models.
Model Structures
- Three-slice inputs performed best among the tested 1-, 3- and 5-slice options, while the 3D model did not deliver the expected overall performance gain over 2D.The 2D model predicts the middle slice from a stack of input slices; RV Dice was better for 3D, and subpixel blocks did not improve performance.Source: Model Structures · p. 7
Data Augmentation
- Light random rotations and scaling improved scores for both models after heavier augmentation yielded sub-par test performance.The authors suggest heavy augmentation might misrepresent heart anatomy.Source: Data Augmentation · pp. 7–8
Optimization Functions
- Dice loss significantly improved both models; the authors report their best Dice scores with this objective and describe it as more robust to pixel-level error optimization.They state that Dice loss better captures spatial context over the image.Source: Optimization Functions · p. 8
5 Conclusion
The paper presents lightweight, modular 2D and 3D CNN pipelines for fully automated cardiac MR segmentation, reporting near-state-of-the-art distance accuracy and convincing clinical accuracy.
5 Conclusion
- The paper introduces fully automated 2D and 3D CNN pipelines for cardiac MR segmentation, with lightweight modular architectures and efficient runtime.Source: §5 · p. 8
- Both models achieved near-state-of-the-art accuracy on distance metrics and convincing performance on clinical accuracy metrics.Source: §5 · p. 8
- Future work would use the segmentation models to learn and classify different cardiac diseases.Source: §5 · p. 8