Abstract
Detecting clinically significant prostate cancer (csPCa) on biparametric MRI (bpMRI) under patient-level weak supervision is challenging because lesion-level annotations are rarely available at scale.This study present a systematic ablation of the Local–Global Multiple Instance Learning (LoGoMIL) architecture introduced by Redekop et al. (2022) on the public PI-CAI 2022 benchmark (1,476 patients), evaluating eleven configurations across six orthogonal design axes: local-only vs local-global context, encoder pretraining strategy (pretrained MedicalNet vs. patch-level pretraining), attention pooling variant in the local branch (plain vs. gated), classifier head (linear vs. MLP), encoder freeze/unfreeze policy, and input modality count (T2W, T2W+ADC, T2W+ADC+HBV). All experiments share a five-fold cross-validation protocol with patient-level AUROC reporting. Two findings stand out. First, adding the global gland-level branch improves the local only baseline significantly. Second, the single largest gain comes from the number of input modalities: adding ADC to T2W improves the T2W only model trained from-scratch by approximately 10 percentage points, and including the high b-value DWI sequence yields the best overall configuration (mean AUROC $0.777 \pm 0.031$).
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/EMERGE_019.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=7Lt8LcCffN
BibTex
@InProceedings{LamHar_Feature_MICCAISAT2026,
author = { Lamini, Harunah AND D’Hondt, Fanny AND Jansen, Bart},
title = { { Feature Extraction Strategies for Clinically Significant Prostate Cancer Detection on Biparametric MRI: A Systematic Ablation of the Local-Global MIL Framework } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17266},
month = {pending},
page = {pending}
}