Enhancing Robustness of Deep Learning to Batch Effects from Multi-site Data for Segmentation of Clinically Significant Prostate Cancer on MRI
Year: 2026
Type: article
Abstract: Deep learning (DL) has shown promise in segmenting clinically significant prostate cancer (csPCa) on MRI. However, batch effects arising from multi-site variations impact their generalizability. This study investigates the impact of training data harmonization, diversity, and few-shot fine-tuning on DL in segmenting csPCa on multi-site MRI. 3 T prostate MRI from N = 1822 patients of four sites (public: D₁, N = 1500; D₂, N = 157; institutional: D₃, N = 47; D₄, N = 118) was leveraged. Bi-parametric MRI (T2-weighted, ADC) was harmonized with csPCa lesions delineated by expert radiologists. nnU-Net DL models for csPCa segmentation on MRI were trained separately (C₁ on D₁; C₂ on D₂) and jointly (C₃ on D₁-D₂). The jointly trained C₃ model was then fine-tuned on D₃ and D₄ using few-shot learning in increments of 50%, 75%, and 100% of train data. Models were primarily evaluated on holdout test sets using sensitivity primarily, in addition to AUC, DSC, precision, and Hausdorff distance (paired t -tests for evaluating significance). Batch effects persisted across datasets despite harmonization. For csPCa segmentation on D 1 /D 2 test sets, C 1 achieved sensitivities of 0.51 ± 0.34/0.11 ± 0.17, C 2 achieved 0.23 ± 0.31/0.17 ± 0.22, while C 3 improved to 0.48 ± 0.36/0.33 ± 0.26. On D 3 /D 4 test sets, C 3 achieved zero-shot sensitivities of 0.19 ± 0.22/0.37 ± 0.35. Few-shot fine-tuning on D 3 and D 4 improved sensitivity to 0.32 ± 0.26/0.54 ± 0.30 on D 3 /D 4 test sets after using 100% of available fine-tuning samples. On subset analyses, lesions > 0.5 cm 3 had consistently higher segmentation performance compared to smaller lesions < 0.5 cm 3 . Diverse multi-site training improves robustness of DL models for csPCa segmentation. Pre-trained models perform modestly on institutional MRI datasets under zero-shot inference, but few-shot fine-tuning on target sites enhances performance.
Institutions: Indiana University Indianapolis, University of Pennsylvania, Georgia Institute of Technology, Emory University, The Wallace H. Coulter Department of Biomedical Engineering, Indiana University School of Medicine, United States Department of Veterans Affairs, Veterans Health Administration, Atlanta VA Medical Center, Indiana University – Purdue University Indianapolis
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $4,690
Funders: U.S. Department of Veterans Affairs, Biomedical Laboratory Research and Development, VA Office of Research and Development, National Institutes of Health, NIH Office of the Director, National Cancer Institute, National Institute of Allergy and Infectious Diseases, National Institute of Diabetes and Digestive and Kidney Diseases, U.S. National Library of Medicine
Awards: G1031957857, G1046651757, G1365000891, G1478154455, G1803346691, Multimodal, Multi-scale Framework for Ethical AI Model Development, G1909130050, G2117823825, G309834616, G3541864657, G3602628895, G5288052009, G5679709768, G5861339792, G6100242255, G7105568102, G8796037806, G8958113727, Novel Radiomics for Predicting Response to Immunotherapy for Lung Cancer, Multi-Consortia Coordinating Center (MC2 Center) for Cancer Biology: Building Interdisciplinary Scientific Communities, Coordinating Impactful Resource Sharing, and Advancing Cancer Research, Quantitative imaging phenotypic classifier for distinguishing radiation effects from tumor recurrence in Glioblastoma, G949296313
OpenAlex ID: W7204229900