Abstract
Accurate segmentation of pancreatic tumors from computed tomography (CT) remains challenging due to low-contrast boundaries, heterogeneous tumor appearance, and extreme foreground–background imbalance. Although transformer-based volumetric segmentation networks have improved organ-level performance, they often exhibit limited lesion-level precision, producing spurious tumor detections that reduce clinical reliability. We propose a precision-oriented extension of Swin UN-ETR that explicitly addresses false-positive suppression through four co-ordinated components: (i) channel-wise contrast normalization at the input, (ii) a multi-scale texture feature extractor at the encoder bottleneck, (iii) a curriculum-based training objective that first promotes tumor recall and subsequently activates a hard-mined false-positive penalty, and (iv) a temperature-scaled softmax inference pipeline with confidence- and size-ranked lesion selection to suppress low-confidence connected components. On the MSD Task07 pancreas benchmark, evaluated on a held-out test set under a strictly controlled data split, the proposed method achieves a tumor Dice Similarity Coefficient (DSC) of 0.452, a lesion-level F1-score of 0.758, a precision of 0.649, and a recall of 0.909, compared with 0.355, 0.620, 0.541, and 0.727 for a plain Swin UNETR baseline trained under identical conditions. At the lesion detection level, false-positive counts are reduced by 21% (34→27) and false-negative counts by 67% (15→5). Post-processing parameters optimized on the validation set further yield a lesion F1 of 0.824 and recall of 0.955, though test-set figures are reported as primary results. These findings demonstrate that curriculum-based precision-oriented training combined with multi-scale texture modeling substantially improves lesion-level recall and specificity in pancreatic tumor segmentation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MLMI_026.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=Et4VvDLMNh
BibTex
@InProceedings{SabMuh_Reducing_MICCAISAT2026,
author = { Sabir, Muhammad AND Hayashi, Yuichiro AND Oda, Masahiro AND Mori, Kensaku},
title = { { Reducing False Positives in Pancreatic Tumor Segmentation via Multi-Scale Texture Enhancement and Curriculum Learning } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17272},
month = {pending},
page = {pending}
}