Deformable MRI Sequence Registration for AI-based Prostate Cancer Diagnosis

TL;DR

The paper tackles misalignment between bpMRI sequences (T2W and DWI) in prostate cancer diagnosis and evaluates whether image registration at inference can improve AI-based csPCa detection. It develops a two-stage registration framework (rigid then deformable) in MeVisLab, aligning ADC/HBV to T2W and applying the deformation to DWI maps within a prostate-restricted domain, guided by a NGF-based similarity and a curvature regularizer. Using the PI-CAI, PCNN, and PROMIS datasets, the study shows deformable registration substantially improves lesion overlap (median Dice around 0.58 versus 0.48) but yields only a small, non-significant AUROC gain (+0.3%, p=0.18) for downstream diagnosis. The authors conclude that improved alignment alone does not guarantee diagnostic gains and advocate for joint, end-to-end development of registration and diagnostic AI systems, including retraining to adapt to registration-induced image changes to optimize patient outcomes.

Abstract

The PI-CAI (Prostate Imaging: Cancer AI) challenge led to expert-level diagnostic algorithms for clinically significant prostate cancer detection. The algorithms receive biparametric MRI scans as input, which consist of T2-weighted and diffusion-weighted scans. These scans can be misaligned due to multiple factors in the scanning process. Image registration can alleviate this issue by predicting the deformation between the sequences. We investigate the effect of image registration on the diagnostic performance of AI-based prostate cancer diagnosis. First, the image registration algorithm, developed in MeVisLab, is analyzed using a dataset with paired lesion annotations. Second, the effect on diagnosis is evaluated by comparing case-level cancer diagnosis performance between using the original dataset, rigidly aligned diffusion-weighted scans, or deformably aligned diffusion-weighted scans. Rigid registration showed no improvement. Deformable registration demonstrated a substantial improvement in lesion overlap (+10% median Dice score) and a positive yet non-significant improvement in diagnostic performance (+0.3% AUROC, p=0.18). Our investigation shows that a substantial improvement in lesion alignment does not directly lead to a significant improvement in diagnostic performance. Qualitative analysis indicated that jointly developing image registration methods and diagnostic AI algorithms could enhance diagnostic accuracy and patient outcomes.

Figures (6)

• Figure 1: Overview of our method. The T2W scan is used as fixed image and the ADC map as moving image to find the displacement field using the registration method. The displacement field is applied to the ADC and HBV maps. The registered and original scans are used as input for the PI-CAI AI system (see \ref{['sec:picai_ai_system']}) to detect clinically significant prostate cancer. The case-level diagnosis performance of the end-to-end pipeline is evaluated and used as a measure of effectiveness.
• Figure 2: This figure demonstrates the impact of synthetic misalignment. (left) the diagnostic performance of the PI-CAI AI system is shown on the PROMIS dataset. When a severe misalignment is introduced, the AUC decreases from 0.793 to 0.720. In the case of extreme misalignment, the AUC further drops to 0.487. (right) shows the T2W and misaligned ADC (top: severe misalignment, bottom: extreme misalignment) images, with prostate gland contour of the T2W scan.
• Figure 3: Quantitative registration results. (left) Distribution of Dice scores between the lesion annotation on the T2W and ADC scans for the original, rigidly aligned, and deformably aligned PCNN datasets. (right) Model performance for the PI-CAI AI system with the original, rigidly aligned and deformably aligned PROMIS datasets.
• Figure 4: Qualitative registration results showing two exemplary cases with prostate gland, lesion annotated on T2, lesion annotated on ADC. In the last two columns, the prediction maps (PM) generated with the original dataset and the deformably aligned dataset are overlayed on the T2W scan.
• Figure 5: Qualitative results on the PROMIS data set. The T2, ADC, and deformably aligned ADC are shown with prostate gland. In the last two columns, the prediction maps (PM) generated with the original dataset and the deformbly aligned dataset are overlayed on the T2W image. The label shows the ISUP grade, where 1 is indolent cancer (negative), and $\geq 2$ is intermediate to high-risk cancer (positive). The first two cases were selected to have the largest prediction increase and decrease for the deformably aligned dataset, compared to the original datasets, for cases with a case-level prediction above 0.3, respectively. The third case was a failure case with the deformably aligned scans.