Journal of Soft Computing and Information Technology

Journal of Soft Computing and Information Technology

Test-Time Augmentation with Multi-Scale Strategies and Test-Time Dropout

Document Type : Persian Original Article

Authors
Emad Asghari
hedieh Sajedi
School of Mathematics, Statistics and Computer Science, College of Science, University of Tehran, Tehran, Iran.
10.22034/jscit.2026.542434.2162
Abstract
Skin lesion segmentation, despite advances in deep learning models, still suffers from performance degradation when faced with challenges such as variations in image quality, lesion heterogeneity, and distribution shifts in test data. Since deep learning models are typically trained once in research environments and, after being deployed in clinical systems, retraining or fine-tuning them with new data is not easily feasible due to the need for labeled data, time, and substantial computational resources this study proposes a lightweight, architecture-free approach to enhance the inference performance of the baseline EGE-UNet model. The proposed method applies two complementary test-time techniques Test-Time Augmentation (TTA) and Test-Time Dropout (TTD) along with their multi-scale variant, as an additional inference layer without modifying the pretrained network. The core idea is to generate multiple augmented or rescaled versions of the input image and execute the pretrained model multiple times with active Dropout, producing more stable and reliable predictions through appropriate averaging. Experiments on the ISIC2017 and ISIC2018 datasets show that combining TTA and TTD improves segmentation accuracy and reduces prediction variance compared to the baseline, while the combining multi-scale version of TTA and TTD yields sharper boundary delineation in ISIC2017; However, there was no improvement in ISIC2018. Because the proposed strategy requires no retraining and can be applied to any segmentation architecture, it represents a practical and low-overhead solution for improving inference robustness in medical imaging applications.
Keywords
Skin lesion segmentation
Test-Time Augmentation
Test-Time Dropout
Multi-Scale Inference
Ensembling
Medical image analysis
  [1]     G. Wang, W. Li, et al., “Aleatoric uncertainty estimation with test-time augmentation for medical image segmentation with convolutional neural networks,” Neurocomputing, vol. 429, pp. 66–73, 2021. DOI: 10.1016/j.neucom.2020.11.103
  [2]     Y. Gal and Z. Ghahramani, “Dropout as a Bayesian approximation: Representing model uncertainty in deep learning,” in Proc. 33rd Int. Conf. Mach. Learn. (ICML), 2016, pp. 1050–1059. DOI: 10.48550/arXiv.1506.02142
  [3]     R. Cao, K. Ma, et al., “EGE-UNet: An efficient group enhanced U-Net for skin lesion segmentation,” Comput. Biol. Med., vol. 142, p. 105197, 2022. DOI: 10.1016/j.compbiomed.2022.105197
  [4]     J. Liang, R. He, and T. Tan, “A comprehensive survey on test-time adaptation under distribution shifts,” Int. J. Comput. Vis., vol. 133, no. 1, pp. 31–64, 2024. DOI: 10.1007/s11263-023-01925-4
  [5]     D. Wang, E. Shelhamer, S. Liu, B. Olshausen, and T. Darrell, “Tent: Fully test-time adaptation by entropy minimization,” in Proc. Int. Conf. Learn. Represent. (ICLR), 2021. DOI: 10.48550/arXiv.2006.10726
  [6]     X. Zhang, Y. Qin, J. Huang, S. Yang, and M. Xu, “LTTA: Lightweight test-time adaptation via updating early layers and uncertainty estimation,” in Adv. Neural Inf. Process. Syst. (NeurIPS), 2024. DOI: 10.48550/arXiv.2311.08252
  [7]     S. Qin, W. Zhang, and Z. Li, “SelfEnsemble for test-time adaptation,” Neurocomputing, vol. 653, p. 131038, Nov. 2025. DOI: 10.1016/j.neucom.2025.131038
  [8]     J. Wang, Y. Liu, C. Gong, and Y. Qiao, “Feature alignment and uniformity for test-time adaptation,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2023, pp. 19452–19461. DOI: 10.1109/CVPR52729.2023.01865
  [9]     J. Yu, H. Liu, and S. Wang, “Train/test-time adaptation with retrieval,” arXiv preprint, 2023. DOI: 10.48550/arXiv.2303.14333
[10]      J. Kang, N. Kim, J. Ok, and S. Kwak, “MemBN: Robust test-time adaptation via batch norm with statistics memory,” in Proc. Eur. Conf. Comput. Vis. (ECCV), 2024. DOI: 10.1007/978-3-031-72664-4_6
[11]      J. Sun, M. Ibrahim, M. Hall, I. Evtimov, Z. M. Mao, C. C. Ferrer, and C. Hazirbas, “VPA: Fully test-time visual prompt adaptation,” in Proc. 31st ACM Int. Conf. Multimedia (MM), 2023, pp. 5796–5806. DOI: 10.1145/3581783.3612293
[12]     International Skin Imaging Collaboration, “ISIC 2017 challenge dataset.” [Online]. Available: https://challenge.isicarchive.com/data/#2017. DOI: N/A
[13]     C. F. Codella et al., “Skin lesion analysis toward melanoma detection: A challenge at the 2017 ISBI,” in Proc. IEEE 15th Int. Symp. Biomed. Imaging (ISBI), Washington, DC, USA, 2018. DOI: 10.1109/ISBI.2018.8363547
[14]     International Skin Imaging Collaboration, “ISIC 2018 challenge dataset.” [Online]. Available: https://challenge.isicarchive.com/data/#2018. DOI: N/A
[15]     N. Codella et al., “Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the ISIC,” arXiv preprint, 2019. DOI: 10.48550/arXiv.1902.03368
[16]     A. Paszke et al., “PyTorch: An imperative style, high-performance deep learning library,” in Adv. Neural Inf. Process. Syst., vol. 32, 2019. DOI: 10.48550/arXiv.1912.01703
[17]     I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” in Proc. Int. Conf. Learn. Represent. (ICLR), 2019. DOI: 10.48550/arXiv.1711.05101
[18]     O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional networks for biomedical image segmentation,” in Proc. MICCAI, 2015, pp. 234–241. DOI: 10.1007/978-3-319-24574-4_28
[19]     X. Gao, M. Zhou, D. Liu, and D. Metaxas, “A multiscale transformer for medical image segmentation,” arXiv preprint, 2022. DOI: 10.48550/arXiv.2203.00131
[20]     Y. Zhang, H. Liu, and Q. Hu, “TransFuse: Fusing transformers and CNNs for medical image segmentation,” in Proc. MICCAI, 2021, pp. 14–24. DOI: 10.1007/978-3-030-87231-1_2
[21]     S. Mehta and M. Rastegari, “Separable self-attention for mobile vision transformers,” arXiv preprint, 2022. DOI: 10.48550/arXiv.2206.02680
[22]     A. Howard et al., “Searching for MobileNetV3,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2019, pp. 1314–1324. DOI: 10.1109/ICCV.2019.00140
[23]     J. M. J. Valanarasu and V. M. Patel, “U-NeXt: MLP-based rapid medical image segmentation network,” arXiv preprint, 2022. DOI: 10.48550/arXiv.2203.04967
[24]     J. Ruan, S. Xiang, et al., “MALUNet: A multi-attention and lightweight U-Net for skin lesion segmentation,” in Proc. IEEE Int. Conf. Bioinf. Biomed. (BIBM), 2022, pp. 1150–1156. DOI: 10.1109/BIBM55620.2022.9995165
[25]     Z. Zhou, M. M. R. Siddiquee, N. Tajbakhsh, and J. Liang, “U-Net++: A nested U-Net architecture for medical image segmentation,” in Deep Learning in Medical Image Analysis, Springer, 2018, pp. 3–11. DOI: 10.1007/978-3-030-00889-5_1
[26]     O. Oktay et al., “Attention U-Net: Learning where to look for the pancreas,” arXiv preprint, 2018. DOI: 10.48550/arXiv.1804.03999
[27]     J. Wei et al., “Shallow attention network for polyp segmentation,” in Proc. MICCAI, 2021, pp. 699–708. DOI: 10.1007/978-3-030-87231-1_64
Journal of Soft Computing and Information Technology
Volume 15, Issue 1
Spring - April
Spring 2026
Pages 17-28