In medical imaging segmentation tasks, patch sampling strategies commonly rely on positive-negative sampling, a method proven effective in frameworks such as nnU-Net and across various publicly available datasets. Existing benchmarks primarily consist of images with limited fields of view or relatively large targets. In contrast, medical image segmentation tasks involving large-field-of-view images containing numerous, variably sized lesions, as for example metastatic lesions in whole-body MRI or malignant melanoma lesions on whole-body 18F-FDG PET/CT, pose distinct challenges. Typically, these images contain several small metastatic lesions that contribute less than 0.01% of the total image volume. We introduce a novel data-centric instance-balanced sampling strategy tailored for these scenarios, significantly improving CPU data loading efficiency, training speed, and downstream segmentation performance. On a representative whole-body dataset it boosts the mean Dice coefficient by 4% (0.66 to 0.70), raises lesion-level sensitivity from 0.63 to 0.70, and lowers false positives per image by 23% (28.5 to 22.0), while reducing data-loading memory consumption to 4% of that required by conventional sampling. Although few open datasets yet share these large-field, lesion-dense characteristics, we anticipate a trend towards adoption of AI in oncological workflows for monitoring metastatic disease, raising the importance of our contribution.
Wuts, J, Ceranka, J, Vandemeulebroucke, J & Lecouvet, F 2026, Instance-Balanced Patch Sampling for Whole-Body Lesion Segmentation. in B Bhattarai, A Rau, R Caramalau, A Reinke, A Nguyen, A Namburete, P Gyawali & D Stoyanov (eds), Data Engineering in Medical Imaging: Third MICCAI Workshop, DEMI 2025. Held in Conjunction with MICCAI 2025. Lecture Notes in Computer Science, Springer, pp. 180-189. https://doi.org/10.1007/978-3-032-08009-7_18
Wuts, J., Ceranka, J., Vandemeulebroucke, J., & Lecouvet, F. (2026). Instance-Balanced Patch Sampling for Whole-Body Lesion Segmentation. In B. Bhattarai, A. Rau, R. Caramalau, A. Reinke, A. Nguyen, A. Namburete, P. Gyawali, & D. Stoyanov (Eds.), Data Engineering in Medical Imaging: Third MICCAI Workshop, DEMI 2025. Held in Conjunction with MICCAI 2025 (pp. 180-189). (Lecture Notes in Computer Science). Springer. https://doi.org/10.1007/978-3-032-08009-7_18
@inproceedings{8055e9c3338a4720b9ab45238c65a352,
title = "Instance-Balanced Patch Sampling for Whole-Body Lesion Segmentation",
abstract = "In medical imaging segmentation tasks, patch sampling strategies commonly rely on positive-negative sampling, a method proven effective in frameworks such as nnU-Net and across various publicly available datasets. Existing benchmarks primarily consist of images with limited fields of view or relatively large targets. In contrast, medical image segmentation tasks involving large-field-of-view images containing numerous, variably sized lesions, as for example metastatic lesions in whole-body MRI or malignant melanoma lesions on whole-body 18F-FDG PET/CT, pose distinct challenges. Typically, these images contain several small metastatic lesions that contribute less than 0.01% of the total image volume. We introduce a novel data-centric instance-balanced sampling strategy tailored for these scenarios, significantly improving CPU data loading efficiency, training speed, and downstream segmentation performance. On a representative whole-body dataset it boosts the mean Dice coefficient by 4% (0.66 to 0.70), raises lesion-level sensitivity from 0.63 to 0.70, and lowers false positives per image by 23% (28.5 to 22.0), while reducing data-loading memory consumption to 4% of that required by conventional sampling. Although few open datasets yet share these large-field, lesion-dense characteristics, we anticipate a trend towards adoption of AI in oncological workflows for monitoring metastatic disease, raising the importance of our contribution.",
author = "Joris Wuts and Jakub Ceranka and Jef Vandemeulebroucke and Fr{\'e}d{\'e}ric Lecouvet",
note = "Publisher Copyright: {\textcopyright} The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.",
year = "2026",
doi = "10.1007/978-3-032-08009-7_18",
language = "English",
isbn = "9783032080080",
series = "Lecture Notes in Computer Science",
publisher = "Springer",
pages = "180--189",
editor = "Binod Bhattarai and Anita Rau and Razvan Caramalau and Annika Reinke and Anh Nguyen and Ana Namburete and Prashnna Gyawali and Danail Stoyanov",
booktitle = "Data Engineering in Medical Imaging",
}