ACDC: Automated Cardiac Diagnosis Challenge

• ACDC is a benchmark for both automatic segmentation of cardiac structures and diagnosis of diseases using cine MRI.
• It involves precise annotations at end-diastole and end-systole for LV, RV, and MYO, paired with diagnostic labels for five conditions.
• ACDC has spurred diverse methodologies from deep learning segmentation models to end-to-end diagnosis pipelines and interpretability studies.

Searching arXiv for foundational and papers on the Automated Cardiac Diagnosis Challenge (ACDC). The Automated Cardiac Diagnosis Challenge (ACDC) is a public benchmark for automatic segmentation of key cardiac structures in cine MRI and automatic diagnosis of cardiac disease from those images. Across the works that use it as a central benchmark, ACDC is consistently defined around 2D short-axis cine cardiac MRI, with annotations for the left ventricle cavity (LV), right ventricle cavity (RV), and myocardium (MYO), and with patient-level diagnostic labels spanning five categories: NOR, DCM, HCM, MINF, and RV/ARV/RVA depending on paper-specific naming conventions (Wolterink et al., 2017, Khened et al., 2018, Janik et al., 2021). It functions both as a segmentation benchmark and as a substrate for end-to-end or staged automated diagnosis pipelines, including feature-based classifiers, radiomics systems, motion- and biomechanics-based methods, multi-task learning systems, and interpretability studies (Isensee et al., 2017, Snaauw et al., 2018, Cetin et al., 2019, Valentin et al., 8 Jul 2025).

1. Benchmark definition and dataset composition

In the ACDC challenge setting, the modality is 4D cine cardiac MRI or, equivalently, 2D short-axis slices over time, with emphasis on end-diastole (ED) and end-systole (ES) for annotation and evaluation (Wolterink et al., 2017, Isensee et al., 2017). The training labels comprise manual segmentations of LV cavity, RV cavity, and myocardium, typically only at ED and ES, plus a patient-level diagnosis label (Wolterink et al., 2017, Isensee et al., 2017, Snaauw et al., 2018).

Several papers restate the canonical cohort structure. One formulation gives 150 exams (150 patients) split into 100 training cases and 50 test cases, with five equally sized groups: NOR, MINF, DCM, HCM, and ARV (Khened et al., 2018). Another formulation describes 150 patients in total, with 100 in the official ACDC training set and 50 in the test set, adopting exactly that split for training and testing (Valentin et al., 8 Jul 2025). A more restricted use appears in interpretability work that focuses on the 100 labeled cases and the challenge’s 70/30 training/development partition rather than the hidden test set (Janik et al., 2021).

The imaging protocol is repeatedly described in compatible terms: short-axis cine MR, SSFP or steady-state free precession, 28–40 time frames per cardiac cycle, slice thickness 5–8 mm or 5–10 mm, and in-plane resolution 1.37–1.68 mm²/pixel (Janik et al., 2021, Khened et al., 2018, Wolterink et al., 2017). Some descriptions also note acquisition on 1.5 T Siemens Avanto and 3.0 T Siemens Trio Tim scanners (Janik et al., 2021).

The five diagnostic groups are tied to clinically defined criteria involving ventricular volumes, ejection fraction, wall motion, myocardial mass, or myocardial thickness (Khened et al., 2018). A plausible implication is that ACDC was designed not merely as a segmentation leaderboard, but as a benchmark linking anatomy, function, and diagnosis.

2. Core tasks: segmentation and diagnosis

ACDC defines two primary tasks: segmentation of LV, RV, and MYO at ED and ES, and cardiac diagnosis into one of the five disease classes (Janik et al., 2021). The segmentation task is commonly formalized as a pixel-wise or voxel-wise labeling problem. One explicit 2D formulation is:

$$X \in \mathbb{R}^{n \times n}, \qquad Y \in \{LV, RV, MYO, bg\}^{n \times n}$$

so that each pixel is assigned to one of the three anatomical structures or background (Janik et al., 2021).

Most ACDC diagnosis pipelines are structured as a sequence: segment the cardiac structures, compute clinical indices such as volumes, ejection fraction, mass, or wall thickness, and then classify the patient into NOR, DCM, HCM, MINF, or RV/ARV/RVA (Janik et al., 2021, Wolterink et al., 2017, Isensee et al., 2017). This staged design is explicit in works that derive LV volume at ED and ES, RV volume at ED and ES, myocardial volume, ejection fraction, and multiple volume ratios from segmentation masks (Wolterink et al., 2017, Isensee et al., 2017).

Several papers give the standard ejection fraction formula:

$EF = \frac{EDV - ESV}{EDV}$

or its percent form (Wolterink et al., 2017, Isensee et al., 2017, Tsai et al., 5 May 2025). In feature-driven ACDC systems, these quantities are then fed to classifiers such as Random Forest, SVM, MLP, Logistic Regression, or ensembles thereof (Wolterink et al., 2017, Isensee et al., 2017, Khened et al., 2018, Valentin et al., 8 Jul 2025).

A recurrent misconception is that ACDC is only a segmentation dataset. The literature repeatedly treats diagnosis as equally central: the challenge is explicitly about combined segmentation plus diagnosis, and many methods are evaluated on both tasks (Wolterink et al., 2017, Khened et al., 2018, Tsai et al., 5 May 2025).

3. Segmentation methodologies developed on ACDC

ACDC has served as a testbed for a wide range of segmentation architectures. Early and influential approaches include 2D fully convolutional networks, U-Net-inspired models, DenseNet-derived FCNs, and dilated CNNs (Wolterink et al., 2017, Khened et al., 2018, Wolterink et al., 2017). A later wave includes directional feature map modules, residual 3D U-Nets, temporal ConvLSTM decoders, foundation-model hybrids, and state-space-model architectures (Cheng et al., 2020, Tsai et al., 5 May 2025, Chen et al., 2020, Huo et al., 22 May 2025).

One ACDC challenge solution employs a 2D fully convolutional network called DFCN-C, combining DenseNet connectivity, residual shortcuts in the decoder, and an Inception-style multi-scale first layer, with about 371k parameters and overall challenge rank second place for segmentation (Khened et al., 2018). Another ACDC pipeline uses a dilated convolutional network with a 131 × 131 receptive field and eight output channels jointly predicting ED and ES classes for LV, RV, MYO, and background (Wolterink et al., 2017).

Temporal structure has also been exploited. A myocardial sequence-segmentation framework combines Res U-net and ConvLSTM, reporting that temporal modeling can improve Dice by up to 2% on ACDC (Chen et al., 2020). A different line of work introduces directional feature maps, learning for each pixel a vector pointing from the nearest boundary into the structure interior, and reports improvement from mean Dice 0.886 to 0.916 and mean Hausdorff distance from 23.009 mm to 6.693 mm on an ACDC train/validation split (Cheng et al., 2020).

More recent methods incorporate larger pretrained or long-range modeling components. SAMba-UNet integrates SAM2, VMamba, and UNet, reporting Dice coefficient of 0.9103 and HD95 boundary error of 1.0859 mm on ACDC, with particular gains for RV and MYO (Huo et al., 22 May 2025). A plausible implication is that ACDC continues to function as a comparative platform across successive architectural paradigms, from compact FCNs to hybrid foundation-model systems.

4. Diagnosis pipelines and clinically derived features

A dominant ACDC design pattern is to convert segmentations into a patient-level feature vector reflecting routine cardiology measurements. Repeated feature families include LV/RV volumes at ED and ES, ejection fractions, myocardial volume or mass, volume ratios, and myocardial wall thickness statistics (Wolterink et al., 2017, Isensee et al., 2017, Tsai et al., 5 May 2025).

One feature-based system computes 14 features: patient height and weight, six volumes, two ejection fractions, and four ratios such as $V_{RV}^{ED} / V_{LV}^{ED}$ and $V_{Myo}^{ES} / V_{LV}^{ES}$ (Wolterink et al., 2017). Another ACDC solution uses 20 features including volumes at ED and ES, LV and RV ejection fractions, LV/RV and Myo/LV volume ratios, and multiple statistics of myocardial wall thickness across slices and phases (Tsai et al., 5 May 2025). The DenseNet-based challenge-winning diagnosis method also derives myocardial wall thickness variation profile features at ED and ES and explicitly uses them in a second-stage expert classifier for MINF vs DCM (Khened et al., 2018).

Radiomics-based diagnosis extends beyond these conventional descriptors. A radiomics study extracts 567 features from ACDC: 3 patient/global features plus 188 features per structure across LV, MYO, and RV, spanning shape-based, first-order intensity, GLCM, GLRLM, and GLSZM categories (Cetin et al., 2019). Sequential forward selection identifies a 10-feature subset that yields 100% accuracy under leave-one-out cross-validation on the 100 ACDC training cases (Cetin et al., 2019). This suggests that ACDC supports not only morphology-function pipelines but also higher-dimensional feature engineering.

End-to-end learning has also been attempted. A multi-task DenseNet/U-Net model jointly learns diagnosis and segmentation from ACDC, reducing classification error from 32% to 22% compared with a diagnosis-only baseline (Snaauw et al., 2018). This suggests that segmentation supervision can regularize diagnosis learning when the dataset is relatively small.

5. Reported performance across representative ACDC systems

The literature reports performance using both segmentation metrics and patient-level diagnostic accuracy. For segmentation, the most common overlap metric is the Dice coefficient:

$\text{Dice}(A,B) = \frac{2|A \cap B|}{|A| + |B|}$

as stated in multiple ACDC works (Janik et al., 2021, Valentin et al., 8 Jul 2025, Cheng et al., 2020). Many studies also report Hausdorff distance, HD95, IoU, or boundary distances (Khened et al., 2018, Huo et al., 22 May 2025, Sander et al., 2020).

The table below summarizes representative concrete results reported on ACDC.

Work	Segmentation result	Diagnosis result
“Automatic Segmentation and Disease Classification Using Cardiac Cine MR Images” (Wolterink et al., 2017)	Average Dice scores 0.94 (LV), 0.88 (RV), 0.87 (myocardium)	91% correct disease category
“Automatic Cardiac Disease Assessment on cine-MRI via Time-Series Segmentation and Domain Specific Features” (Isensee et al., 2017)	Test Dice 0.950 (LVC), 0.923 (RVC), 0.911 (LVM)	92% on test set
“Fully Convolutional Multi-scale Residual DenseNets...” (Khened et al., 2018)	second place for segmentation in ACDC-2017	first place with 100% accuracy
“Biomechanics-Informed Neural Network for Explainable Cardiovascular Disease Identification” (Valentin et al., 8 Jul 2025)	Dice 0.945 (LV), 0.908 (RV), 0.905 (MYO)	98% train, 100% test
“IntelliCardiac” (Tsai et al., 5 May 2025)	Average Dice 92.56%	98% accuracy
“SAMba-UNet” (Huo et al., 22 May 2025)	Dice 0.9103, HD95 1.0859 mm	Not a diagnosis classifier

These results are not directly interchangeable because protocols differ. Some studies use the official 100/50 ACDC split (Valentin et al., 8 Jul 2025, Tsai et al., 5 May 2025); some use internal cross-validation on the 100 labeled cases (Wolterink et al., 2017, Isensee et al., 2017, Cetin et al., 2019); and some focus only on the 70/30 training/development partition or the public training set (Janik et al., 2021, Sander et al., 2020). A plausible implication is that direct leaderboard-style ranking from paper tables alone can be misleading unless the evaluation protocol is matched.

6. Interpretability, robustness, and methodological extensions

A significant later development in ACDC research is the shift from pure performance reporting to explainability, quality control, and physically informed modeling. One interpretability study trains a 2D U-Net for LV/RV/MYO segmentation on ACDC and probes it using D-TCAV, a concept-based method that automatically discovers latent concepts from super-pixels and assigns TCAV scores per pathology class (Janik et al., 2021). The core finding is that concepts learned by a segmentation-only model are mostly anatomical, not pathology-specific, with average maximum–minimum pathology difference across clusters of 3.4% and a notable outlier in cluster 56 (Janik et al., 2021). This suggests that anatomy-trained representations do not automatically become disease-discriminative.

Another line of work combines segmentation with uncertainty estimation and local failure detection. Using ACDC, one study trains Bayesian versions of DN, DRN, and U-net, derives uncertainty maps using entropy and MC-dropout, and then trains a second CNN to identify local segmentation failures (Sander et al., 2020). Simulated manual correction of detected failure regions yields statistically significant improvements in Dice and Hausdorff distance, and reduces editing time from about 20 minutes to less than 2 minutes per patient in a manual-correction experiment (Sander et al., 2020). This suggests that ACDC can support not just automatic segmentation, but semi-automatic reliability-aware workflows.

Biomechanics-informed approaches move beyond static anatomy. A 2025 method builds a cascaded CNN-based 3D registration framework with Neo-Hookean regularization on ACDC, estimates local deformation, and derives mechanical features such as voxelwise effective $\mu(p)$ and $\kappa(p)$ for diagnosis (Valentin et al., 8 Jul 2025). This explicitly links cardiac motion and tissue mechanics to ACDC disease classes. A plausible implication is that ACDC has become a benchmark for integrating image analysis with physical modeling, not only for segmentation accuracy.

7. Role of ACDC in the cardiac MRI research landscape

Across the cited literature, ACDC serves several distinct functions. First, it is a segmentation benchmark for LV, RV, and MYO in cine MRI, supporting comparisons across CNNs, transformers, state-space models, and hybrid systems (Khened et al., 2018, Cheng et al., 2020, Huo et al., 22 May 2025). Second, it is an automated diagnosis benchmark in which segmentation-derived features, radiomics, or learned representations are mapped to the five canonical classes (Wolterink et al., 2017, Isensee et al., 2017, Cetin et al., 2019, Snaauw et al., 2018). Third, it is a methodological sandbox for interpretability, uncertainty estimation, motion extraction, registration, and biomechanics (Janik et al., 2021, Sander et al., 2020, Upendra et al., 2021, Valentin et al., 8 Jul 2025).

A recurring theme is that ACDC’s disease taxonomy is sufficiently structured to reward clinically motivated features. Many high-performing methods rely on measurements that directly reflect the class definitions: ventricular dilatation for DCM, myocardial thickening for HCM, regional thinning or altered thickness heterogeneity for MINF, and RV enlargement or dysfunction for ARV/RVA (Wolterink et al., 2017, Khened et al., 2018, Tsai et al., 5 May 2025). This has preserved the relevance of interpretable, feature-driven pipelines even as end-to-end models have become more capable.

At the same time, several works note limitations. Some approaches are developed and validated on ACDC only and explicitly identify external validation and domain generalization as future work (Tsai et al., 5 May 2025, Tamoor et al., 2024). Some are constrained by the fact that only ED and ES are manually annotated, even though the underlying data are full cine sequences (Isensee et al., 2017, Snaauw et al., 2018). Others note that ACDC leaderboard-style metrics do not capture interpretability, reliability, or clinical workflow integration (Janik et al., 2021, Sander et al., 2020).

Taken together, the corpus suggests that ACDC occupies a foundational place in automated cardiac MRI analysis: it anchors research on segmentation, diagnosis, explainability, and functional modeling within a single, shared benchmark. A plausible implication is that its enduring value lies not only in its leaderboard history but in the way it has structured an entire research program around the relationship between cardiac anatomy, function, pathology, and algorithmic transparency.