Maximizing T2-Only Prostate Cancer Localization from Expected Diffusion Weighted Imaging

Weixi Yi; Yipei Wang; Wen Yan; Hanyuan Zhang; Natasha Thorley; Alexander Ng; Shonit Punwani; Fernando Bianco

cs.CVApr 1, 2026

Maximizing T2-Only Prostate Cancer Localization from Expected Diffusion Weighted Imaging

Authors: Weixi Yi, Yipei Wang, Wen Yan, Hanyuan Zhang, Natasha Thorley, Alexander Ng, Shonit Punwani, Fernando Bianco, +5 more

Organizations: the UCL Hawkes Institute and the Department of Medical Physics and Biomedical Engineering, University College London, London, UK · the Centre for Medical Imaging, University College London, London, UK · the Division of Surgery & Interventional Science, University College London, London, UK · Urological Research Network, Miami Lakes, Florida, USA · the Centre for Bioengineering, School of Engineering and Materials Science and Digital Environment Research Institute, Queen Mary University of London, London, UK

Abstract

Multiparametric MRI is increasingly recommended as a first-line noninvasive approach to detect and localize prostate cancer, requiring at minimum diffusion-weighted (DWI) and T2-weighted (T2w) MR sequences. Early machine learning attempts using only T2w images have shown promising diagnostic performance in segmenting radiologist-annotated lesions. Such uni-modal T2-only approaches deliver substantial clinical benefits by reducing costs and expertise required to acquire other sequences. This work investigates an arguably more challenging application using only T2w at inference, but to localize individual cancers based on independent histopathology labels. We formulate DWI images as a latent modality (readily available during training) to classify cancer presence at local Barzell zones, given only T2w images as input. In the resulting expectation-maximization algorithm, a latent modality generator (implemented using a flow matching-based generative model) approximates the latent DWI image posterior distribution in the E-steps, while in M-steps a cancer localizer is simultaneously optimized with the generative model to maximize the expected likelihood of cancer presence. The proposed approach provides a novel theoretical framework for learning from a privileged DWI modality, yielding superior cancer localization performance compared to approaches that lack training DWI images or existing frameworks for privileged learning and incomplete modalities. The proposed T2-only methods perform competitively or better than baseline methods using multiple input sequences (e.g., improving the patient-level F1 score by 14.4% and zone-level QWK by 5.3% over the T2w+DWI baseline). We present quantitative evaluations using internal and external datasets from 4,133 prostate cancer patients with histopathology-verified labels.

Figures & tables

Fig. 1: Prostate cancer localization paradigms. Unlike standard mpMRI requiring multi-sequence inputs, our proposed Generalized EM leverages DWI strictly as a latent modality during training. This enables high-accuracy inference using only T2w images, overcoming the limitations of standard T2 models. Red denotes ISUP $\geq$ 3.

Fig. 2: The left shows the proposed GEM framework with two components: a latent modality generator $f_{G}$ generating latent images $\hat{x}_{z}$ from observed images $x_{o}$ , and a cancer localizer $f_{L}$ localizing lesions using $x_{o}$ and $\hat{x}_{z}$ . Biopsy-derived Barzell-zone labels $y$ are colored orange and red for ISUP grades 2 and $\geq$ 3, with others labeled negative. The right shows GEM’s graphical model: the latent variable $\hat{x}_{z}$ is generated from $x_{o}$ and $\theta$ (E-step), while $\theta$ is updated using $\hat{x}_{z}$ and data $\mathcal{T}$ (M-step). $r$ is omitted for clarity.

Fig. 3: Assignment of Barzell zones to the PZ or TZ. (a) T2w. (b) DWI. (c) PZ/TZ contours overlaid on Barzell zones. (d) Final assignment based on the majority overlap rule.

Methods	Training	Inference	QWK $\uparrow$	AUC $\uparrow$	Macro F1 $\uparrow$	MCC $\uparrow$
Baseline(T2w,DWI)	T2w,DWI	T2w,DWI	$25.34\pm 3.58$	$68.94\pm 0.81$	$33.68\pm 3.35$	$16.12\pm 2.03$
Baseline(T2w,ADC)	T2w,ADC	T2w,ADC	$10.94\pm 0.86$	$64.47\pm 1.50$	$28.71\pm 0.94$	$12.47\pm 2.16$
Baseline(3M)	T2w,DWI,ADC	T2w,DWI,ADC	$31.36\pm 3.42$	$71.80\pm 1.76$	$35.95\pm 3.07$	$19.02\pm 2.85$
Baseline(DWI)	DWI	DWI	$27.82\pm 1.31$	$69.45\pm 0.67$	$36.29\pm 1.41$	$17.62\pm 2.09$
Baseline(ADC)	ADC	ADC	$20.05\pm 3.80$	$66.56\pm 2.78$	$32.96\pm 2.88$	$16.69\pm 1.10$
ModDrop [ 35 ]	T2w,DWI	T2w	$11.42\pm 2.02$	$62.88\pm 0.86$	$27.67\pm 1.18$	$8.21\pm 0.94$

TABLE I: Barzell zone-level ISUP group prediction. Best in bold ; second best underlined .

Methods	Tertiary Definition				Primary Definition				Secondary Definition
Methods	Spe@Sen80% $\uparrow$	Spe@Sen90% $\uparrow$	Sen@Spe80% $\uparrow$	Sen@Spe90% $\uparrow$	Spe@Sen80% $\uparrow$	Spe@Sen90% $\uparrow$	Sen@Spe80% $\uparrow$	Sen@Spe90% $\uparrow$	Spe@Sen80% $\uparrow$	Spe@Sen90% $\uparrow$	Sen@Spe80% $\uparrow$	Sen@Spe90% $\uparrow$
Baseline(T2w,DWI)	$44.74\pm 1.49$	$29.58\pm 1.76$	$43.74\pm 0.79$	$25.98\pm 2.23$	$46.73\pm 2.08$	$31.03\pm 1.65$	$45.87\pm 2.34$	$31.35\pm 2.23$	$50.95\pm 5.65$	$32.93\pm 11.05$	$57.69\pm 0.00$	$43.59\pm 8.67$
Baseline(T2w,ADC)	$44.82\pm 7.23$	$30.24\pm 5.99$	$32.60\pm 3.43$	$14.43\pm 1.24$	$41.23\pm 9.66$	$24.22\pm 6.43$	$34.19\pm 1.44$	$15.68\pm 2.73$	$34.71\pm 1.66$	$19.62\pm 5.21$	$34.62\pm 7.69$	$16.03\pm 1.11$
Baseline(3M)	$51.91\pm 2.55$	$35.99\pm 1.19$	$46.39\pm 4.50$	$29.41\pm 3.12$	$56.19\pm 2.32$	$38.31\pm 3.77$	$56.37\pm 3.17$	$40.76\pm 4.76$	$52.65\pm 9.77$	$37.74\pm 12.19$	$53.85\pm 13.46$	$43.46\pm 16.08$
Baseline(DWI)	$45.70\pm 2.52$	$30.27\pm 2.33$	$42.65\pm 1.82$	$28.45\pm 0.87$	$49.59\pm 2.79$	$33.68\pm 2.63$	$51.98\pm 1.31$	$38.45\pm 1.03$	$48.26\pm 4.23$	$33.32\pm 6.68$	$55.13\pm 2.94$	$40.38\pm 13.46$
Baseline(ADC)	$45.70\pm 1.07$	$30.87\pm 2.21$	$42.65\pm 5.63$	$20.82\pm 5.12$	$44.70\pm 2.39$	$29.22\pm 3.05$	$37.72\pm 8.46$	$21.09\pm 7.19$	$41.24\pm 9.98$	$29.82\pm 6.73$	$41.67\pm 16.58$	$29.49\pm 16.36$

TABLE II: Barzell zone-level csPCa detection. Best in bold ; second best underlined .

Methods	Peripheral Zone				Transition Zone
Methods	F1 $\uparrow$	AUC $\uparrow$	Spec@Sen80% $\uparrow$	Sen@Spe80% $\uparrow$	F1 $\uparrow$	AUC $\uparrow$	Spec@Sen80% $\uparrow$	Sen@Spe80% $\uparrow$
Baseline(T2w,DWI)	$58.24\pm 8.34$	$64.16\pm 5.87$	$37.78\pm 7.79$	$41.67\pm 19.25$	$49.75\pm 5.48$	$62.31\pm 1.65$	$35.97\pm 8.44$	$36.36\pm 5.25$
Baseline(T2w,ADC)	$64.21\pm 1.92$	$64.65\pm 8.25$	$47.56\pm 18.68$	$27.78\pm 12.11$	$54.79\pm 6.12$	$59.45\pm 4.44$	$43.33\pm 9.44$	$30.30\pm 3.03$
Baseline(3M)	$63.44\pm 3.39$	$70.66\pm 3.58$	$45.63\pm 14.64$	$49.07\pm 8.93$	$61.54\pm 1.47$	$69.76\pm 3.90$	$44.86\pm 3.13$	$49.29\pm 7.67$
Baseline(DWI)	$57.17\pm 5.29$	$63.33\pm 0.33$	$39.11\pm 2.70$	$40.74\pm 5.78$	$57.82\pm 2.89$	$68.54\pm 4.71$	$46.25\pm 4.91$	$40.61\pm 11.70$
Baseline(ADC)	$62.21\pm 1.26$	$69.55\pm 3.23$	$56.15\pm 1.85$	$39.81\pm 10.52$	$59.47\pm 2.68$	$61.20\pm 4.45$	$39.17\pm 10.47$	$31.72\pm 2.99$

TABLE III: Region-level (PZ/TZ) csPCa detection under the primary definition. Best in bold ; second best underlined .

Methods	PROMIS				PI-CAI
Methods	F1 $\uparrow$	AUC $\uparrow$	Spec@Sen80% $\uparrow$	Sen@Spe80% $\uparrow$	F1 $\uparrow$	AUC $\uparrow$	Spec@Sen80% $\uparrow$	Sen@Spe80% $\uparrow$
Baseline(T2w,DWI)	$65.39\pm 3.94$	$64.06\pm 1.94$	$33.33\pm 6.78$	$38.81\pm 9.83$	$51.15\pm 1.60$	$71.84\pm 1.56$	$50.67\pm 1.88$	$48.25\pm 2.89$
Baseline(T2w,ADC)	$72.18\pm 1.90$	$62.41\pm 4.04$	$42.74\pm 7.83$	$27.41\pm 7.33$	$45.49\pm 0.41$	$63.58\pm 2.09$	$36.66\pm 2.72$	$35.06\pm 4.72$
Baseline(3M)	$68.98\pm 2.28$	$69.67\pm 5.37$	$49.57\pm 7.40$	$46.67\pm 13.33$	$49.99\pm 1.08$	$72.34\pm 1.38$	$51.05\pm 2.76$	$50.90\pm 3.50$
Baseline(DWI)	$67.52\pm 5.93$	$62.94\pm 0.34$	$35.90\pm 2.56$	$39.56\pm 5.98$	$49.29\pm 3.92$	$69.93\pm 2.18$	$45.09\pm 3.11$	$46.29\pm 2.36$
Baseline(ADC)	$72.71\pm 2.85$	$66.12\pm 3.44$	$44.44\pm 7.40$	$32.00\pm 7.75$	$46.08\pm 0.15$	$64.51\pm 1.74$	$38.56\pm 3.01$	$36.12\pm 2.13$

TABLE IV: Patient-level csPCa detection under the primary definition. Best in bold ; second best underlined .

CRF	Modality Generator	GEM	QWK $\uparrow$	AUC $\uparrow$	Macro F1 $\uparrow$	MCC $\uparrow$
			$9.21\pm 2.21$	$59.74\pm 1.67$	$25.56\pm 3.91$	$9.06\pm 3.13$
✓			$9.54\pm 6.17$	$60.75\pm 1.13$	$25.88\pm 3.55$	$9.90\pm 5.06$
	✓		$17.55\pm 2.94$	$66.35\pm 0.82$	$30.87\pm 3.17$	$13.53\pm 1.85$
✓	✓		$19.99\pm 3.78$	$66.21\pm 1.72$	$31.82\pm 1.99$	$14.59\pm 2.14$
✓	✓	✓	$\mathbf{26.68\pm 0.65}$	$\mathbf{70.30\pm 2.08}$	$\mathbf{34.20\pm 2.92}$	$\mathbf{18.06\pm 3.14}$

TABLE V: Ablation study of different components on ISUP group prediction performance.

Fig. 4: Samples of T2w images, corresponding real DWI, and DWI synthesized by pretrained LFM, Ours(LDM), and Ours(LFM). The rightmost column shows the corresponding Barzell zone labels by ISUP grade, where white denotes ISUP $=$ 0, yellow denotes ISUP $=$ 1, orange denotes ISUP $=$ 2, and red denotes ISUP $\geq$ 3.

$\alpha$	QWK $\uparrow$	AUC $\uparrow$	Macro F1 $\uparrow$	MCC $\uparrow$
0.05	$24.67\pm 1.62$	$68.84\pm 0.17$	$33.69\pm 1.12$	$17.68\pm 0.28$
0.1	$\mathbf{26.68\pm 0.65}$	$\mathbf{70.30\pm 2.08}$	$\mathbf{34.20\pm 2.92}$	$\mathbf{18.06\pm 3.14}$
0.2	$24.44\pm 0.94$	$68.83\pm 0.21$	$33.82\pm 1.23$	$17.80\pm 0.25$
0.5	$24.11\pm 0.67$	$68.84\pm 0.22$	$33.52\pm 0.91$	$17.29\pm 0.66$

TABLE VI: Ablation study of different

\alpha

on ISUP group prediction performance.

Explore similar work

Mar 19, 2026

Interpretable Prostate Cancer Detection using a Small Cohort of MRI Images

Prostate cancer is a leading cause of mortality in men, yet interpretation of T2-weighted prostate MRI remains challenging due to subtle and heterogeneous lesions. We developed an interpretable framework for automatic cancer detection using a small dataset of 162 T2-weighted images (102 cancer, 60 normal), addressing data scarcity through transfer learning and augmentation. We performed a comprehensive comparison of Vision Transformers (ViT, Swin), CNNs (ResNet18), and classical methods (Logistic Regression, SVM, HOG+SVM). Transfer-learned ResNet18 achieved the best performance (90.9% accuracy, 95.2% sensitivity, AUC 0.905) with only 11M parameters, while Vision Transformers showed lower performance despite substantially higher complexity. Notably, HOG+SVM achieved comparable accuracy (AUC 0.917), highlighting the effectiveness of handcrafted features in small datasets. Unlike state-of-the-art approaches relying on biparametric MRI (T2+DWI) and large cohorts, our method achieves competitive performance using only T2-weighted images, reducing acquisition complexity and computational cost. In a reader study of 22 cases, five radiologists achieved a mean sensitivity of 67.5% (Fleiss Kappa = 0.524), compared to 95.2% for the AI model, suggesting potential for AI-assisted screening to reduce missed cancers and improve consistency. Code and data are publicly available.

Vahid Monfared, Mohammad Hadi Gharib, Ali Sabri +4

Department of Computer Science, Boston University Metropolitan College, Boston, MA, USA · MD, Clinical Fellow, Department of Radiology, McMaster University, Ontario, Canada · Department of Radiology, School of Medicine, 5th Azar Hospital, Radiology, Golestan University of Medical Sciences, Gorgan, Iran +6

Medical Imaging Medical Image Classification Explainable Medical Image Analysis

Ask PDF

Jul 13, 2026

Diffusion MRI preprocessing affects ADC estimation and automatic PI-RADS v2.1 classification in bi-parametric prostate MRI

Diffusion-weighted imaging (DWI) is acquired as part of bi-parametric prostate MRI, but suffers from artifacts that degrade downstream quantitative and diagnostic performance. While DWI preprocessing is standard in brain imaging, its adoption in prostate imaging remains limited and lacks standardized pipelines. This study investigated the effect of different DWI preprocessing strategies on apparent diffusion coefficient (ADC) estimation and automatic Prostate Imaging Reporting and Data System (PI-RADS) classification. 268 cases were derived from the fastMRI prostate cohort by sequentially applying denoising, Gibbs-ringing correction, and diffeomorphic registration for susceptibility distortion correction. ADC maps were compared using linear least squares (LLS) and iteratively-weighted LLS (IWLLS). A 3-class DenseNet classifier was trained to predict PI-RADS scores from multi-channel MRI inputs. ADC analysis revealed statistically significant differences across preprocessing pipelines, with LLS and IWLLS producing numerically equivalent maps. Linear relationships between ADC values were preserved across most datasets (PCC ~0.99), while distortion correction realigned DWI to T2w anatomy and altered ADC values accordingly (PCC ~0.90). Classification showed the best AUROC and sensitivity for high-risk PI-RADS classes in the fully processed dataset. False-negative analysis revealed this dataset produced the least overconfident incorrect predictions on high-risk classes, which is a desirable property for clinical triage. DWI preprocessing, particularly distortion correction, enhances both ADC map quality and the predictive power of deep learning models for PI-RADS classification, supporting the need for optimized preprocessing pipelines in prostate MRI.

Christos Kanakis, Mathias Perslev, Tim Schakel +4

Center for Image Sciences, University Medical Center Utrecht, Utrecht, The Netherlands · Cerebriu, Copenhagen, Denmark · Department of Radiotherapy, University Medical Center Utrecht, Utrecht, The Netherlands +2

Medical Image Classification Magnetic Resonance Imaging

Ask PDF

Jul 19, 2026

Histopathological Spectrum-Guided Prostate Stratification via Segmentation-Assisted Diagnostic Transformer

Prostate cancer diagnosis with multiparametric MRI (mpMRI) is commonly based on PI-RADS assessment or binary classification, which suffer from subjectivity and fail to capture clinically relevant pathological heterogeneity. To address this limitation, we construct a Prostate Cancer Histopathology Spectrum Dataset (PCa-HSD) and formulate a clinically meaningful four-class classification task, addressing the underrepresentation of benign lesions that are easily confounded with prostate cancer in existing datasets. We propose Language-guided Segmentation-assisted Diagnostic Transformer model (LSDT), which leverages zero-shot segmentation to provide anatomical priors and performs effective multi-modal slice fusion for classification. Our proposed method consistently improves accuracy across backbones, achieving the best average accuracy of 0.633 and JointRecall of 0.768 in five-fold cross-validation on a cohort of 344 patients. These results demonstrate that integrating pathology supervision and anatomical priors significantly enhances fine-grained prostate MRI classification and provides a more clinically relevant paradigm for risk stratification. Code will be made publicly available in a future revision.

Leyang Li, Lihua Chen, Huangang Hu +7

College of Artificial Intelligence, Nankai University, Tianjin 300350, China · Tianjin First Central Hospital, Tianjin 300382, China · School of Electronics and Information Engineering, Tiangong University, Tianjin 300387, China +3

Medical Image Classification Magnetic Resonance Imaging

Ask PDF