Abstract
Radiologist assessment of prostate mpMRI can miss a non-trivial fraction of clinically significant prostate cancer (csPCa), leaving uncertainty about where suspicious tissue may reside when no targetable lesion is identified. Most deep learning approaches are trained to reproduce radiologist-defined targets, such as PI-RADS–scored radiologist-identified lesions, or their biopsy-verified subset, which may encourage models to over-focus on radiologically dominant cues. We propose a novel training strategy to detect “radiologically invisible” tissue by formulating a dense patch-level classification task supervised by biopsy-derived Barzell zone labels. To reduce the model’s over-reliance on obvious signals, we introduce virtual lesion excision: a strategy that suppresses radiologist-marked lesions and their peri-lesional tissue during training. This encourages the network to prioritise residual contextual cues and subtle imaging signatures that remain after suppression. Experiments on a biopsy-blind cohort show that lesion suppression improves patch-level discrimination over a standard baseline trained without suppression, based on template-guided saturation biopsy as reference standard. In a workload-constrained setting, where only a small number of candidate patches can be presented per patient, an intermediate amount of suppression yields the strongest operating performance: at 1/2/5 false-positive patches per patient, sensitivity increases to 0.244 ± 0.100, 0.348 ± 0.087, and 0.515 ± 0.091 versus a baseline of 0.183 ± 0.106, 0.285 ± 0.156, and 0.428 ± 0.180. These findings suggest that lesion-suppression training is a promising and simple intervention, providing potential decision support in cases where routine mpMRI interpretation does not identify a clear target.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/CaPTion_037.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/profile?id=~Maruf_Talukdar1
BibTex
@InProceedings{TalMar_Why_MICCAISAT2026,
author = { Talukdar, Maruf AND Wang, Yipei AND Thorley, Natasha AND Huang, Shiqi AND Kasivisvanathan, Veeru AND Punwani, Shonit AND Emberton, Mark AND Hu, Yipeng},
title = { { Why Background is Important: Virtual Lesion Excision for Detecting Radiologically “Invisible” Prostate Cancer } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17278},
month = {pending},
page = {pending}
}