ProstAtlasDiff: Prostate cancer detection on MRI using Diffusion Probabilistic Models guided by population spatial cancer atlases

Abstract

Magnetic Resonance Imaging (MRI) is increasingly being used to detect prostate cancer, yet its interpretation can be challenging due to subtle differences between benign and cancerous tissue. Recently, Denoising Diffusion Probabilistic Models (DDPMs) have shown great utility for medical image segmentation, modeling the process as noise removal in standard Gaussian distributions. In this study, we further enhance DDPMs by introducing the knowledge that the occurrence of cancer varies across the prostate (e.g., ∼70% of prostate cancers occur in the peripheral zone). We quantify such heterogeneity with a registration pipeline to calculate voxel-level cancer distribution mean and variances. Our proposed approach, ProstAtlasDiff, relies on DDPMs that use the cancer atlas to model noise removal and segment cancer on MRI. We trained and evaluated the performance of ProstAtlasDiff in detecting clinically significant cancer in a multi-institution multi-scanner dataset, and compared it with alternative models. In a lesion-level evaluation, ProstAtlasDiff achieved statistically significantly higher accuracy (0.91 vs. 0.85, p<0.001), specificity (0.91 vs. 0.84, p<0.001), positive predictive value (PPV, 0.50 vs. 0.35, p<0.001), compared to alternative models. ProstAtlasDiff also offers more accurate cancer outlines, achieving a higher Dice Coefficient (0.33 vs. 0.31, p<0.01). Furthermore, we evaluated ProstAtlasDiff in an independent cohort of 91 patients who underwent radical prostatectomy to compare its performance to that of radiologists, relative to whole-mount histopathology ground truth. ProstAtlasDiff detected 16% (15 lesions out of 93) more clinically significant cancers compared to radiologists (sensitivity: 0.90 vs. 0.75, p<0.01), and was comparable in terms of ROC-AUC, PR-AUC, PPV, accuracy, and Dice coefficient (p≥0.05). Furthermore, we evaluated ProstAtlasDiff in a second independent cohort of 537 subjects and observed that ProsAtlasDiff outperformed alternative approaches. These results suggest that ProstAltasDiff has the potential to assist in localizing cancer for biopsy guidance and treatment planning.

Keywords: Denoising Diffusion Probabilistic Models; Magnetic Resonance Imaging; Population atlases; Prostate cancer detection.

Fig. 1.

Atlas construction process. We constructed the atlas by pairwise registration of every subject T2w image to the reference, guided by the prostate boundary (showing the prostate glands in beige and light green, csPCa annotations in orange and dark green). The iterative registration process includes Step (1) Centering and Scaling, Step (2) Affine Registration and Step (3) Deformable registration, and each followed is by updating the reference atlas. The process allows the construction of the mean atlas, $A_{\mathit{\text{cancer}}}^{\mathit{\text{mean}}}$, and the variance atlas, $A_{\mathit{\text{cancer}}}^{\mathit{\text{variance}}}$, for the csPCa labels in the training set; The green to purple color represents an increase in the occurrence of cancer.

Fig. 2.

Pipeline for the proposed ProstAtlasDiff model. $A_{\mathit{\text{cancer}}}^{\mathit{\text{mean}}}$ and $A_{\mathit{\text{cancer}}}^{\mathit{\text{variance}}}$ are registered to the original image, resulting in case-specific mean and variance atlases $A_{\mu }$ (Blue) and $A_{\sigma }$ (Pink). The proposed ProstAtlasDiff design is based on diffusion models using cancer atlases as both prior distribution and additional input. The noisy segmentation features, MRI features, and atlas features are extracted with separate encoding branches. Feature fusion is performed using intermediate features $F_{seg},F_{img}$, and $F_{\mathit{\text{atlas}}}$ after each downsampling layer (feature fusion module in lilac), to enhance the encoding of cancer features in the prediction encoder. Features from the last layer of the three encoders are concatenated together and fed into the decoder. Skip connections are added between the prediction encoder and the decoder at the same resolution. The decoder then outputs the predicted noise to be removed. Consequently, a corresponding one-step denoised $x_{t-1}$ was output.

Fig. 3.

csPCa prediction comparison among different models. Left to right: T2, ADC, DWI with csPCa outlined in green; nnUNet, nnUNet-A, DecNet, DecNet-A, MedSegDiff, MedSegDiff-A, ProstAtlasDiff, with csPCa prediction in purple, and ground truth lesion outlined in green.

Fig. 4.

Comparison of ProstAtlasDiff predictions, Radiologist annotation, and pathologist labels for a patient that underwent radical prostatectomy and the whole mount was registered to MRI. The MRI is shown from apex (left) to base (right). Pathologist outlines of cancer are shown in yellow (ISUP Grade Group ≥ 2) and blue (ISUP Grade Group = 1). Radiologist annotations (green) and ProstAtlasDiff predictions (purple) are overlaid on T2w in the top row.

Fig. 5.

Receiver operating characteristic curves for lesion-level prediction of different models for the biopsy cohort $C_{Bx}$. The diagonal dashed line indicates a model with random prediction that corresponds to ROC-AUC = 0.5.