Differentiable Deconvolution for Improved Stroke Perfusion Analysis

Ezequiel de la Rosa, David Robben, Diana Maria Sima, Jan S. Kirschke, Bjoern H. Menze

TL;DR

This paper introduces the first end-to-end trainable framework for stroke perfusion analysis that integrates a differentiable SVD deconvolution layer. By making the classical physics-based deconvolution step differentiable, the authors allow a neural network to automatically learn the optimal Arterial Input Function (AIF) that maximizes the accuracy of ischemic core lesion segmentation.

What Problem Does This Paper Address?

In acute ischemic stroke, doctors use perfusion imaging to distinguish between "penumbra" (salvageable tissue) and the "core" (irreversibly damaged tissue). The standard mathematical tool for this is Singular Value Decomposition (SVD) deconvolution, which converts the passage of a contrast agent through the brain into meaningful maps like Cerebral Blood Flow (CBF).

However, SVD deconvolution is incredibly sensitive to a specific input: the Arterial Input Function (AIF)—the concentration curve of the contrast agent in a feeding artery. Currently, AIF selection is often done manually by radiologists, which is slow, highly variable, and prone to bias. While automatic heuristic methods exist, they don't necessarily pick the AIF that results in the most accurate clinical diagnosis. This paper aims to automate AIF selection by optimizing it directly for the final goal: accurate lesion segmentation.

Methods

The authors propose a hybrid pipeline that combines deep learning with classical signal processing:

  1. AIF Generating CNN: A regression network (using 3D and 2D convolutions) takes 4D perfusion data as input and outputs a 1D vector representing the AIF ($c_{art}(t)$).
  2. Differentiable SVD Deconvolution: This is the core innovation. Instead of treating deconvolution as a "black box" post-processing step, they implement it as a differentiable layer using Volterra discretization and Tikhonov regularization. This allows gradients from the segmentation loss to flow back through the math of the deconvolution and into the AIF-generating CNN.
  3. Lesion Mapping: The resulting relative CBF (rCBF) map is passed through a sigmoid activation to produce a final lesion probability map ($y_{pred}$).
The proposed differentiable SVD deconvolution pipeline.

By training the whole system end-to-end using a Soft-Dice loss against ground-truth lesions (from Diffusion Weighted Imaging, or DWI), the network learns to "regress" the specific AIF shape that yields the best segmentation results.

Datasets

The study utilized the ISLES18 (Ischemic Stroke Lesion Segmentation 2018) database.

  • Scope: 94 training cases containing 4D CT perfusion series.
  • Ground Truth: Expert-delineated core lesions obtained from DWI, which is the gold standard for identifying irreversible damage.
  • Diversity: The dataset is multi-scanner and multi-center, providing a robust test for generalization.

Key Insights and Ideas

  • Optimization for Task, Not Imitation: Unlike previous AI attempts that tried to mimic how a human picks an AIF, this model picks an AIF based on what actually improves the final segmentation. The "best" AIF for a mathematical model might not look exactly like the one a human picks in a large artery.
  • Physics-Informed Deep Learning: By embedding the SVD deconvolution directly into the network, the model is constrained by the underlying physics of perfusion. It isn't just "guessing" a lesion map; it's using a well-validated physiological model as an internal layer.
  • Removing Rater Bias: Because the network requires no manual AIF annotations for training, it eliminates the subjectivity and variability inherent in human expert performance.

Key Results

The method was evaluated using 5-fold cross-validation and compared against both manual expert selection and a state-of-the-art FDA-cleared software (icobrain cva).

  • Segmentation Accuracy: The proposed method achieved a Dice score of 0.359, slightly outperforming the manual expert (0.353) and the automatic software (0.351), though the differences were not statistically significant.
  • Discriminative Power: The rCBF maps generated by this method had a higher Area Under the ROC Curve (AUC of 0.868) compared to the expert and automatic methods (both 0.856).
  • Efficiency: The entire automated process takes approximately 1.15 seconds per case, compared to roughly one minute for manual expert annotation.
Qualitative comparison of rCBF maps and resulting segmentations.

Limitations and Future Work

  • Sample Size: The study was limited to 94 subjects. The authors expect that a larger, more diverse dataset would further improve the regression performance.
  • Parameter Scope: Currently, the model only optimizes for the rCBF map. In clinical practice, the Tmax (time to maximum) map is also vital.
  • Differentiability Challenges: Optimizing for Tmax is harder because it involves an argmax operation, which is not naturally differentiable. The authors identify finding a differentiable substitute for argmax as a key area for future research.
  • Volumetric Bias: The use of Soft-Dice loss led to a slight volumetric overestimation (bias), a known characteristic of that specific loss function.

Summary updated

Abstract

Perfusion imaging is the current gold standard for acute ischemic stroke analysis. It allows quantification of the salvageable and non-salvageable tissue regions (penumbra and core areas respectively). In clinical settings, the singular value decomposition (SVD) deconvolution is one of the most accepted and used approaches for generating interpretable and physically meaningful maps. Though this method has been widely validated in experimental and clinical settings, it might produce suboptimal results because the chosen inputs to the model cannot guarantee optimal performance. For the most critical input, the arterial input function (AIF), it is still controversial how and where it should be chosen even though the method is very sensitive to this input. In this work we propose an AIF selection approach that is optimized for maximal core lesion segmentation performance. The AIF is regressed by a neural network optimized through a differentiable SVD deconvolution, aiming to maximize core lesion segmentation agreement with ground truth data. To our knowledge, this is the first work exploiting a differentiable deconvolution model with neural networks. We show that our approach is able to generate AIFs without any manual annotation, and hence avoiding manual rater’s influences. The method achieves manual expert performance in the ISLES18 dataset. We conclude that the methodology opens new possibilities for improving perfusion imaging quantification with deep neural networks.
BibTeX
@inproceedings{rosa2020differentiable,
  author = {Ezequiel de la Rosa and David Robben and Diana Maria Sima and Jan S. Kirschke and Bjoern H. Menze},
  title = {{Differentiable Deconvolution for Improved Stroke Perfusion Analysis}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2020},
  url = {https://openreview.net/forum?id=1c9qdKWz5g},
}

Loading related content...

Figure