Deformable MRI Sequence Registration for AI-based Prostate Cancer Diagnosis

Alessa Hering, Sarah de Boer, Anindo Saha, Jasper J. Twilt, Mattias P. Heinrich, Derya Yakar, Maarten de Rooij, Henkjan Huisman, Joeran S. BosmaarXiv 2024 · arXiv

TL;DR

While synthetic misalignment severely degrades the performance of AI models detecting prostate cancer on biparametric MRI, automated deformable image registration significantly improves physical lesion alignment (+10% median Dice score). Surprisingly, this substantial improvement in alignment does not translate into a statistically significant boost in downstream diagnostic accuracy (+0.3% AUROC), revealing that current AI models are remarkably robust to minor spatial realignments.

What Problem Does This Paper Address?

AI models for prostate cancer diagnosis typically rely on biparametric MRI (bpMRI), which combines structural T2-weighted (T2W) scans with functional diffusion-weighted imaging (DWI). Because these sequences are acquired sequentially, patient movement or internal physiological changes (like bladder filling) frequently cause spatial misalignment between them.

Most state-of-the-art diagnostic AI algorithms use an "early fusion" strategy, stacking these image modalities directly on top of each other. When sequences are misaligned, the AI is forced to combine features from different spatial locations, which can lead to misdiagnoses. While manual alignment by experts fixes this, it is too labor-intensive for clinical deployment. This paper investigates whether automated image registration can solve the misalignment problem and, crucially, whether better alignment actually improves an AI's diagnostic accuracy.

Methods

The authors developed a two-step automated image registration pipeline:

  1. Rigid Registration: Aligns the global orientation using a Normalized Gradient Field (NGF) distance measure to focus on matching image gradients and edges while suppressing noise.
  2. Deformable Registration: Uses a matrix-free deformable registration approach with a second-order curvature regularizer to ensure smooth, physically plausible deformations. To maximize accuracy, this step is restricted strictly to the prostate gland, utilizing an automated prostate segmentation tool.

The calculated transformations are applied to the DWI maps to align them with the fixed T2W scan. The aligned images are then fed into the PI-CAI AI system—an ensemble of the top five deep learning algorithms from the Prostate Imaging: Cancer AI challenge—without any retraining.

Fig. 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 Section 2.3) 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.

Datasets

The study evaluated the pipeline across three distinct datasets:

  • PI-CAI: 10,207 bpMRI cases used originally to develop and train the top-performing diagnostic AI models.
  • PCNN (Prostaat Centrum Noord-Nederland): An independent subset of 473 cases containing manual, voxel-level lesion annotations on both T2W and ADC modalities, used specifically to evaluate physical registration performance.
  • PROMIS: An external validation dataset consisting of 546 cases from 11 UK hospitals, used to evaluate the downstream diagnostic impact of the registration pipeline.

Key Insights and Ideas

  • The Alignment Paradox: Improving physical image alignment does not automatically guarantee better AI performance. Deformable registration successfully brought the modalities into closer alignment, yet the downstream diagnostic AI showed negligible improvement.
  • AI Robustness: Qualitative analysis revealed that diagnostic AI models are surprisingly resilient to minor misalignments. Unless a spatial shift accidentally misrepresents a lesion into an entirely incorrect anatomical zone (e.g., shifting a benign nodule from the transition zone to the peripheral zone), the model can still classify the case correctly.
  • The Hidden Cost of Resampling: Deformable registration requires interpolating and resampling pixel values. This process can introduce subtle artifacts or smooth out critical diagnostic features (like diffusion restriction values in small lesions), potentially offsetting the benefits of better spatial alignment.

Key Results

  • Impact of Misalignment: Introducing synthetic severe misalignment caused the diagnostic AI's Area Under the ROC curve (AUROC) to drop from 0.793 to 0.720. Extreme misalignment degraded it to a near-random 0.487, proving that alignment is critical.
Impact of synthetic misalignment on diagnostic performance
  • Registration Performance: On the PCNN dataset, deformable registration increased the median Dice score of lesion overlap to 0.58, up from 0.48 in the original unaligned dataset. Rigid registration showed no improvement (0.47).
  • Downstream Diagnostic Performance: On the external PROMIS dataset, applying deformable registration yielded a minor, statistically non-significant diagnostic improvement of +0.3% AUROC (moving from 0.793 to 0.796, $p=0.18$).
Quantitative registration and diagnostic results

Limitations and Future Work

A primary limitation is that the diagnostic AI models were evaluated "off-the-shelf" and were not retrained on the newly registered images. Because of this, the models could not adapt to the subtle image variations and smoothing introduced by the registration's resampling step.

The authors conclude that image registration shouldn't just be an isolated preprocessing step. Instead, future work should focus on jointly developing and training image registration methods alongside diagnostic AI algorithms within an end-to-end framework to truly enhance clinical accuracy and patient outcomes.

Summary updated

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.
BibTeX
@misc{hering2024deformable,
  author = {Alessa Hering and Sarah de Boer and Anindo Saha and Jasper J. Twilt and Mattias P. Heinrich and Derya Yakar and Maarten de Rooij and Henkjan Huisman and Joeran S. Bosma},
  title = {{Deformable MRI Sequence Registration for AI-based Prostate Cancer Diagnosis}},
  year = {2024},
  eprint = {2404.09666},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2404.09666},
}

Loading related content...

Figure