VentrEX: An Anatomically Guided Deep Learning Pipeline for Ventricular Segmentation in Cine Cardiac MRI

Figures

• 

  Figure 1. Overview of the VentrEX pipeline from preprocessing and VentrEX-Seg to PM/T standardization and evaluation. ACDC [14] was used for model development and held-out testing, whereas Sunnybrook/LVSC [15] and MM-WHS [16] were used for zero-shot external validation. Blue horizontal arrows indicate the main workflow from the input data through VentrEX-Seg to PM/T standardization. Green arrows denote feature extraction and regression-based threshold prediction, the pink downward arrow connects the pipeline outputs to the evaluation and analysis stage, and gray or black arrows indicate processing or transformation steps within individual stages.

• 

  Figure 2. VentrEX-Seg architecture. 3D encoder-decoder with parallel channel-spatial attention at each scale and a Transformer bottleneck; decoder concatenates skips and a 1 × 1 × 1 head outputs BG, RV, MYO, and LV.

• 

  Figure 3. Patient-level Dice and Precision distributions across models. The upper row shows Dice coefficients, and the lower row shows Precision for the RV, myocardium, and LV. Boxplots represent the patient-level distributions for U-Net, CBAM, nnU-Net, and VentrEX-Seg. Pairwise Wilcoxon signed-rank comparisons are overlaid; (p < 0.05) indicates statistical significance.

• 

  Figure 4. Bland–Altman analysis of LV volumes at ED and ES for the 50-patient held-out ACDC test cohort. Blue circles represent ED observations, and orange triangles represent ES observations. Differences were defined as prediction minus reference. Dashed lines indicate phase-specific mean biases, and dotted lines indicate the corresponding 95% limits of agreement calculated as bias ±1.96 SD.

• 

  Figure 5. Bland–Altman analysis of RV volumes at ED and ES for the 50-patient held-out ACDC test cohort. Blue circles represent ED observations and orange triangles represent ES observations. Differences were defined as prediction minus reference. Dashed lines indicate phase-specific mean biases, and dotted lines indicate the corresponding 95% limits of agreement calculated as bias ±1.96 SD.

• 

  Figure 6. Sunnybrook LV (pooled). External LV volumes exhibit a larger negative bias (underestimation) and wider limits, reflecting cross-dataset differences in contour conventions and acquisition protocols.

• 

  Figure 7. MM-WHS RV (pooled). External RV shows a comparable negative bias and wider limits.

• 

  Figure 8. Perturbation sensitivity by layer group. Violation scores quantify performance degradation after masking attention-identified regions. Encoder D3 and the bottleneck (BL) are most sensitive for LV/MYO, whereas decoder stages show progressively stabilized saliency.

• 

  Figure 9. 3D attention maps (AGC) across classes and layer groups. Attention-weighted activations localize to anatomical borders and are refined from encoder (D3) to bottleneck (BL) to decoder (E2). Darker regions denote higher attention weights.

• 

  Figure 10. Representative 2D overlays on ACDC. Qualitative comparison among U-Net, CBAM, nnU-Net, VentrEX-Seg, and VentrEX. Green contours denote reference; red contours denote predictions. VentrEX reduces papillary-muscle and trabecular “leaks” and yields crisper endocardial borders across ED/ES.

• 

  Figure 11. LV 3D reconstructions across ED/ES (ACDC). Columns show ground truth, direct predictions, and anatomical predictions after PM/T standardization. Removing papillary/trabecular signal from the LV cavity reduces systematic overestimation and produces anatomically consistent endocardial surfaces.

• 

  Figure 12. External 3D ventricular surfaces. Results are shown for (a) MM-WHS and (b) Sunnybrook. The first and third columns show the direct VentrEX-Seg outputs, whereas the second and fourth columns show the final VentrEX outputs after PM/T standardization. The standardized outputs exhibit improved basal/apical continuity and fewer spurious bridges.