Abstract

Prototypical Networks operate by embedding both support and query samples into a common feature space and then represent-ing each class with the mean vector of its support embeddings. Yet, the inherent complexity of medical imagery pose significant challenges for isolating features that are both precise and depend-able. Consequently, constructing effective prototypes in this do-main demands not only sophisticated preprocessing and more pow-erful embedding architectures, but also deliberate refinement of feature representations. In this context, most important and rep-resentative feature map selection is critical. We introduce Selec-tive Feature Representation in Prototypical Networks, a lightweight yet effective enhancement to prototype-based few-shot learning. Proposed approach explicitly refines support embeddings by rank-ing and selecting the top feature maps for each class, leveraging an ensemble of channel-wise statistics—Global Average Pooling, Max Pooling, and Variance. Built on a compact CONV4 back-bone, proposed method outperforms much larger state-of-the-art models on two medical benchmarks: achieving 67.18% (1-shot) and 78.20% (5-shot) on Derm7pt skin-lesion classification, and 63.39% (1-shot), 77.17% (5-shot), and 83.06% (10-shot) on Blood-MNIST pathology classification. These gains demonstrate that tar-geted feature-map selection significantly improves prototype qual-ity and generalization with minimal complexity, offering a practical solution for resource-constrained clinical applications.

References

• Andrea Acevedo, Anna Merino, Santiago Alferez,´ Angel Molina, Laura Boldu,´ and Jose´ Rodellar. A dataset of micro-scopic peripheral blood cell images for development of auto-matic recognition systems. Data in Brief, 30:105474, 2020.
• Arijit Bera, Subhransu Kumar, and Ashwin Machanavajjhala. Exploring medical image mining using deep convolutional networks. In SIAM International Conference on Data Mining, pages 543–551, 2018.
• Runyang Cao, Yu Zhang, Song Bai, and Philip Torr. Comet: Concept-based prototypical networks for few-shot learning. In International Joint Conference on Artificial Intelligence, pages 2108–2114, 2021.
• Yujia Chen, Zhenguo Li, and Ruijia Wang. Feature-level meta-learning with attention for few-shot classification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 512–520, 2022.
• Andre Esteva, Brett Kuprel, Roberto A. Novoa, Justin Ko, Susan M. Swetter, Helen M. Blau, and Sebastian Thrun. Dermatologist-level classification of skin cancer with deep neural networks. In Nature, pages 115–118, 2017.
• J. Kawahara et al. Seven-point checklist and skin lesion clas-sification. IEEE JBHI, 2019.
• Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International Conference on Machine Learning, pages 1126–1135, 2017.
• Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, Franc¸ois Laviolette, Mario Marc-hand, and Victor Lempitsky. Domain-adversarial training of neural networks. In International Conference on Machine Learning, pages 1180–1189, 2016.
• Xia He, Xue Yuan, Kai Zhang, Hu Li, and Wei Liu. Robust re-weighting prototypical networks for few-shot classifica-tion. In Proceedings of the 6th International Conference on Robotics and Artificial Intelligence (ICRAI ’20), pages 125– 132, Singapore, 2020. ACM.
• Timothy M. Hospedales, Antreas Antoniou, Paul Mica-helis, and Amos Storkey. Meta-learning in neural net-works: A survey. In International Conference on Learning Representations, 2020.
• Jieneng Hu and et al. Feature disentanglement and filter-ing for few-shot learning. In Proceedings of the International Conference on Learning Representations (ICLR), 2023.
• Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Im-agenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems, pages 1097–1105, 2012.
• Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. In Proceedings of the 28th International Conference on Neural Information Processing Systems, pages 1–9, 2015.
• Qi Li, Hao Zhang, and Wei Wang. Learning to select in-formative features for few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11835–11844, 2023.
• Geert Litjens, Thijs Kooi, Babak Ehteshami Bejnordi, Arnaud A.A. Setio, Francesco Ciompi, Mohsen Ghafoorian, Jeroen A.W. M. van der Laak, Bram van Ginneken, and Clara I. Sanchez´. A survey on deep learning in medical image anal-ysis. In Medical Image Computing and Computer-Assisted Intervention – MICCAI, pages 295–307, 2017.
• Xiaolong Liu, Yifan Wang, and Lei Zhang. Weighted proto-typical networks for few-shot learning. In AAAI Conference on Artificial Intelligence, pages 9319–9326, 2019.
• Yun Liu, Geert Litjens, and Bram van Ginneken. Deep learn-ing for content-based medical image retrieval: A performance study. In 2016 IEEE International Conference on Image Processing (ICIP), pages 1275–1279, 2016.
• Mingsheng Long, Yue Cao, Jianmin Wang, and Michael I. Jordan. Learning transferable features with deep adaptation networks. In International Conference on Machine Learning, pages 97–105, 2015.
• Xu Luo, Hao Wu, Ji Zhang, Lianli Gao, Jing Xu, and Jingkuan Song. A closer look at few-shot classification again. In International Conference on Machine Learning, pages 23103–23123. PMLR, 2023.
• Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In ICLR Workshop, 2017.
• Jun Shu, Qi Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zong-ben Xu, and Deyu Meng. Meta-weight-net: Learning an ex-plicit mapping for sample weighting. In Advances in Neural Information Processing Systems, NeurIPS 2019, pages 1917– 1928, 2019.
• Jake Snell, Kevin Swersky, and Richard Zemel. Prototypi-cal networks for few-shot learning. In Advances in Neural Information Processing Systems, pages 4077–4087, 2017.
• Koichi Suzuki, Katsuyuki Mori, and Kunio Doi. Medical im-age sorting and retrieval by combining shape and texture fea-tures. In SPIE Medical Imaging, pages 978–985, 2016.
• Oriol Vinyals, Charles Blundell, Tim Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one-shot learning. In Advances in Neural Information Processing Systems, pages 3630–3638, 2016.
• Yongqiang Wang, Yikang Li, Kai Chen, Shiji Song, and Ying Huang. A survey on few-shot learning. In CVPR Workshops, pages 2866–2875, 2020.
• Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? In Advances in Neural Information Processing Systems, pages 3320–3328, 2014.
• Zhongzheng Yue, Qianru Sun, and Qi Yu. Compact fea-ture learning for prototypical few-shot classification. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 10454–10463, 2023.