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.
Introduction
In the management of acute ischemic stroke, perfusion imaging is essential for distinguishing between the "penumbra"—potentially salvageable tissue—and the "core"—irreversibly damaged tissue. The standard mathematical framework for this quantification is Singular Value Decomposition (SVD) deconvolution. This process transforms the temporal passage of a contrast agent through the brain into clinically actionable maps, such as Cerebral Blood Flow (CBF).
A significant challenge in this workflow is the sensitivity of SVD deconvolution to the Arterial Input Function (AIF), which represents the contrast agent's concentration curve in a feeding artery. Traditionally, AIF selection is performed manually by radiologists, a process that is time-consuming, subject to high inter-rater variability, and prone to bias. While automated heuristic methods exist, they often fail to optimize for the ultimate clinical goal: accurate lesion segmentation. This paper introduces an end-to-end trainable framework that integrates a differentiable SVD deconvolution layer, allowing a neural network to automatically learn the optimal AIF for ischemic core lesion segmentation.
Methodology
The proposed pipeline bridges the gap between deep learning and classical signal processing through a hybrid architecture:
- AIF Generating CNN: A regression network utilizing 3D and 2D convolutions processes 4D perfusion data to output a 1D vector representing the AIF ($c_{art}(t)$).
- Differentiable SVD Deconvolution: The core innovation is the implementation of the deconvolution step as a differentiable layer. By employing Volterra discretization and Tikhonov regularization, the authors enable gradients from the segmentation loss to propagate back through the mathematical deconvolution process into the AIF-generating CNN.
- Lesion Mapping: The resulting relative CBF (rCBF) map is transformed via a sigmoid activation into a final lesion probability map ($y_{pred}$).
The entire system is trained using a Soft-Dice loss against ground-truth core lesions derived from Diffusion Weighted Imaging (DWI), the gold standard for identifying irreversible damage. This "task-specific" optimization ensures the network learns the AIF shape that yields the most accurate clinical results, rather than simply mimicking human selection.
Experimental Setup
The study utilized the ISLES18 (Ischemic Stroke Lesion Segmentation 2018) database, comprising 94 training cases with 4D CT perfusion series. This multi-scanner, multi-center dataset provides a robust environment for testing the model's generalization capabilities. Expert-delineated core lesions from DWI served as the ground truth for training and evaluation.
Results and Discussion
The method was evaluated using 5-fold cross-validation and compared against manual expert selection and a state-of-the-art FDA-cleared software solution.
- Segmentation Performance: The proposed method achieved a Dice score of 0.359, slightly outperforming both manual experts (0.353) and automated software (0.351).
- Discriminative Power: The rCBF maps generated by the differentiable model demonstrated superior discriminative ability, with an Area Under the ROC Curve (AUC) of 0.868, compared to 0.856 for both expert and heuristic methods.
- Operational Efficiency: The automated pipeline processes a case in approximately 1.15 seconds, a significant improvement over the roughly one minute required for manual expert annotation.
By embedding the SVD deconvolution directly into the network, the model remains constrained by the underlying physics of perfusion. This physics-informed approach eliminates rater bias and subjectivity, as the network requires no manual AIF annotations during training.
Conclusion
This work demonstrates that differentiable deconvolution layers can effectively combine the interpretability of classical physiological models with the predictive power of deep learning. While the current model focuses on rCBF maps, future work will explore differentiable substitutes for operations like argmax to incorporate other vital parameters like Tmax (time to maximum). Additionally, scaling the approach to larger datasets and addressing the volumetric overestimation bias inherent in Soft-Dice loss remain key areas for development.
Summary updated