Answers are grounded in this paper and cite their sources.
Hybrid CNN--Transformer classifiers can appear advantageous when compared only with fixed or untuned CNN baselines, making it difficult to separate the effect of global-context refinement from hyperparameter optimization and imbalance handling. This study evaluates a compact EfficientNet-B0--Transformer against a matched-budget optimized CNN-only comparator using 3,064 contrast-enhanced T1-weighted MRI slices from 233 patients. All model development used patient-disjoint training, validation, and held-out test partitions. The CNN-only and hybrid pipelines used focal loss and adaptive minority-class augmentation and were independently optimized with 40 Optuna tree-structured Parzen estimator trials; the test set remained inaccessible during selection. On the held-out set, optimized CNN and hybrid models achieved Macro F1 scores of 0.9096 and 0.9341, respectively, with patient Macro F1 of 0.8956 and 0.9327. Across three independent training seeds, hybrid image-level Macro F1 was 0.9249 ± 0.0187 versus 0.9018 ± 0.0090 for the optimized CNN. However, patient-wise five-fold robustness results were close (0.9355 ± 0.0383 versus 0.9342 ± 0.0455), and the paired patient-level McNemar test was not significant (p=0.5). The hybrid required only 9.5% more parameters, with modest latency and memory overhead. The results support compact self-attention as an incremental refinement, while showing that its benefit should be judged against equally optimized CNN baselines and multiple robustness checks.
No takes yet. Share an insight, caveat, or question.
Hardani et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: