IntroductionDeep convolutional neural networks learn rich feature representations; however, the final classification head may not fully align with this feature space after end-to-end training, leading to underutilization of discriminative information. To address this limitation, we propose a lightweight post-hoc redundant head mechanism that improves feature-space adaptation without modifying or retraining the backbone network.MethodsThe proposed approach introduces an additional shallow classification head trained from scratch on frozen feature representations augmented with frozen class logits. This design preserves training stability while enabling broader learning-rate exploration within an expanded signal space. The inclusion of frozen logits provides class-aware priors that stabilize optimization and support improved adaptation within the frozen feature manifold, allowing recovery of residual discriminative information with negligible computational overhead. The method was evaluated using a DenseNet-121 backbone enhanced with a Group–Squeeze-Excitation–Shuffle (GSSE) feature block on the HAM10000 skin lesion dataset through seven-class training and clinically relevant binary evaluation.ResultsThe proposed method improved binary classification accuracy from 91.06% to 92.42% (+1.36 percentage points) and macro-averaged F1-score from 0.8588 to 0.8748 (+0.0160). It also outperformed the strongest baseline accuracy of 91.11% by 1.31 percentage points, while introducing only negligible computational overhead.DiscussionThese findings demonstrate that lightweight post-hoc redundant head learning with logit-aware feature augmentation provides a stable, computationally efficient, and practical strategy for enhancing feature-space adaptation and improving medical image classification performance without requiring modification or retraining of the backbone network.