<p>Deploying deep learning in Cardiac MRI faces a critical trade-off between the diagnostic precision of Vision Transformers (ViTs) and the computational efficiency required for clinical workflows. To address this, we propose a novel Difficulty-Aware Relational Distillation Framework for cardiac pathology classification. We distill complex representational knowledge from a high-capacity Hybrid Swin-CNN Teacher into a highly optimized EfficientNetV2-L Student. Our optimization strategy integrates three core components: Relational Manifold Alignment (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">${L}_{Rel}$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi>L</mi> <mrow> <mi mathvariant="italic">Rel</mi> </mrow> </msub> </math></EquationSource> </InlineEquation>) for structural consistency, Focal Distillation Loss (<InlineEquation ID="IEq2"> <EquationSource Format="TEX">${L}_{Focal}$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi>L</mi> <mrow> <mi mathvariant="italic">Focal</mi> </mrow> </msub> </math></EquationSource> </InlineEquation>) to prioritize challenging samples, and Dynamic Temperature Modulation (<InlineEquation ID="IEq3"> <EquationSource Format="TEX">${\tau}_{i}$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi>τ</mi> <mi>i</mi> </msub> </math></EquationSource> </InlineEquation>) to adaptively scale distillation intensity based on the Teacher’s confidence. Evaluated on 6000 cardiac MRI images via a rigorous patient-level stratified fivefold cross-validation, the framework achieved a state-of-the-art mean Accuracy of 89.85% and a Matthews Correlation Coefficient (MCC) of <InlineEquation ID="IEq4"> <EquationSource Format="TEX">$0.8120$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>0.8120</mn> </mrow> </math></EquationSource> </InlineEquation>. Statistical analysis (Wilcoxon signed-rank test, <InlineEquation ID="IEq5"> <EquationSource Format="TEX">$p<0.05$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>p</mi> <mo><</mo> <mn>0.05</mn> </mrow> </math></EquationSource> </InlineEquation>) confirms that our model significantly outperforms parameter-heavy baselines while delivering a 4.25× reduction in computational complexity, enabling generalizable and efficient high-throughput clinical diagnostics.</p>

Difficulty-aware feature distillation for efficient cardiac MRI classification

期刊论文
发表日期：

2026年6月15日

查看全文

Jafar Abdollahi,
Babak Nouri-Moghaddam,
Aminreza Mohajerzadeh,
Abbas Mirzaei,
Nahideh Derakhshanfard

摘要

Deploying deep learning in Cardiac MRI faces a critical trade-off between the diagnostic precision of Vision Transformers (ViTs) and the computational efficiency required for clinical workflows. To address this, we propose a novel Difficulty-Aware Relational Distillation Framework for cardiac pathology classification. We distill complex representational knowledge from a high-capacity Hybrid Swin-CNN Teacher into a highly optimized EfficientNetV2-L Student. Our optimization strategy integrates three core components: Relational Manifold Alignment ( ${L}_{Rel}$ $L_{Rel}$ ) for structural consistency, Focal Distillation Loss ( ${L}_{Focal}$ $L_{Focal}$ ) to prioritize challenging samples, and Dynamic Temperature Modulation ( ${\tau}_{i}$ $τ_{i}$ ) to adaptively scale distillation intensity based on the Teacher’s confidence. Evaluated on 6000 cardiac MRI images via a rigorous patient-level stratified fivefold cross-validation, the framework achieved a state-of-the-art mean Accuracy of 89.85% and a Matthews Correlation Coefficient (MCC) of $0.8120$ $0.8120$ . Statistical analysis (Wilcoxon signed-rank test, $p<0.05$ $p < 0.05$ ) confirms that our model significantly outperforms parameter-heavy baselines while delivering a 4.25× reduction in computational complexity, enabling generalizable and efficient high-throughput clinical diagnostics.