Self-Supervised Few-Shot Learning for Ischemic Stroke Lesion Segmentation

Luca Tomasetti, Stine Hansen, Mahdieh Khanmohammadi, Kjersti Engan, Liv Jorunn Høllesli, Kathinka Dæhli Kurz, Michael Kampffmeyer

TL;DR

This paper introduces a novel self-supervised few-shot learning approach for segmenting ischemic stroke lesions (ISLs) in brain CT scans. By leveraging color-coded parametric maps (PMs) derived from CT perfusion scans, their method significantly improves segmentation accuracy with very limited labeled data, achieving an average Dice score of 0.58 with just one annotated sample.

What Problem Does This Paper Address?

Accurate segmentation of ischemic stroke lesions is crucial for effective diagnosis and treatment planning. However, current deep learning methods for this task typically require vast amounts of manually annotated data, which is expensive and time-consuming to obtain in the medical domain. This scarcity of labeled data hinders the development of robust models, especially for rare cases. Furthermore, manual annotations can sometimes be imprecise, leading to suboptimal training. This paper aims to overcome these limitations by developing a few-shot learning approach that requires minimal labeled data, specifically focusing on ISL segmentation.

Methods

The core of this work is a novel self-supervised training mechanism integrated into a few-shot segmentation framework called ADNet. Unlike previous self-supervised methods that generate "pseudolabels" from raw image data (which can be problematic for ISLs due to their subtle differences from healthy tissue), this approach leverages domain knowledge by extracting supervoxels from color-coded parametric maps (PMs) derived from CT perfusion (CTP) scans. These PMs (including CBF, CBV, TTP, and TMax) highlight tissue changes and provide more informative regions for supervoxel generation.

The process involves:

  1. Pre-processing: Standard steps to extract brain tissue and enhance contrast from CTP scans.
  2. Supervoxel Generation: A modified 3D Felzenszwalb's algorithm is used to generate supervoxels from stacked PMs. This algorithm is chosen for its ability to produce diverse supervoxels.
  3. Self-Supervised Training: Training episodes are constructed using unlabeled CTP scans. A random supervoxel from the PMs acts as the foreground pseudolabel. Two slices containing this supervoxel are sampled as support and query pairs. Random transformations are applied to one of the pairs to encourage invariance.
  4. ADNet Architecture: The support and query slices are fed into a shared feature encoder (ResNet-101). A foreground prototype is computed from the support features and used to segment the query volume by calculating the negative cosine similarity between the prototype and query features.
  5. Inference: During inference, a single labeled 2D+time slice from a support patient is used to guide the segmentation of query patient volumes slice by slice.

Datasets

The study utilized a dataset of 152 CTP scans from distinct patients. These patients were categorized into three groups:

  • Large Vessel Occlusion (LVO): 77 patients
  • Non-Large Vessel Occlusion (Non-LVO): 60 patients
  • Without Ischemic Stroke (WIS): 15 patients (included for validation and to assess performance on negative cases)

The data was split into training (89 patients), validation (30 patients), and testing (33 patients) sets. The CTP scans and their corresponding PMs were obtained using Siemens Healthineers' "syngo.via" software. Manual annotations were performed by two expert neuroradiologists.

Key Insights and Ideas

  • Leveraging Domain Knowledge is Crucial: The paper strongly argues that using domain-specific information (like PMs) for generating pseudolabels is significantly more effective than relying on raw image data for ISL segmentation. PMs capture subtle physiological changes that are key indicators of stroke.
  • Parametric Maps as Richer Pseudolabels: Supervoxels derived from PMs provide more informative and accurate targets for self-supervised learning compared to supervoxels from raw CTP data, especially for ISLs which can be subtle and heterogeneous.
  • Few-Shot Learning for Medical Imaging: The success of this approach highlights the potential of few-shot learning, particularly self-supervised variants, to address the data scarcity problem in medical image analysis.
  • Improved Robustness: The proposed method shows better performance and more stable segmentations compared to baselines, indicating increased robustness to variations in ISL appearance and patient data.

Key Results

  • Quantitative Performance: The proposed method achieved an average Dice score of 0.58 for ISL segmentation using only a single annotated patient slice. This significantly outperformed baseline methods.
  • Dataset Specificity: Training on the CTP-LVO dataset yielded the best overall performance, suggesting that a more homogeneous training set focused on a specific patient group can lead to better understanding of ISL characteristics.
  • Comparison to Supervised Methods: While the self-supervised approach did not reach the performance of a fully supervised method that uses the entire labeled dataset (e.g., Tomasetti et al. [12] with a Dice score of 0.83-0.85), it achieved encouraging results with vastly less labeled data.
  • Qualitative Results: Visualizations in Figure 4 show that the proposed method's predictions, while not perfect, more closely resemble the ground truth ISL areas compared to baseline methods, which often exhibited under- or over-segmentation.

Limitations and Future Work

  • Performance Gap with Supervised Methods: The paper acknowledges that the achieved performance is not as high as fully supervised methods that utilize extensive labeled data.
  • Non-LVO Patient Segmentation: The method struggled with segmenting ISLs in Non-LVO patients, likely due to the smaller and more subtle nature of these lesions.
  • Future Work: The authors suggest that further research is needed to validate these results and explore more advanced few-shot learning techniques for ISL segmentation. Expanding the dataset and investigating more sophisticated methods for handling less distinct lesions are potential future directions.

Summary updated

Abstract

Precise ischemic lesion segmentation plays an essential role in improving diagnosis and treatment planning for ischemic stroke, one of the prevalent diseases with the highest mortality rate. While numerous deep neural network approaches have recently been proposed to tackle this problem, these methods require large amounts of annotated regions during training, which can be impractical in the medical domain where annotated data is scarce. As a remedy, we present a prototypical few-shot segmentation approach for ischemic lesion segmentation using only one annotated sample during training. The proposed approach leverages a novel self-supervised training mechanism that is tailored to the task of ischemic stroke lesion segmentation by exploiting color-coded parametric maps generated from Computed Tomography Perfusion scans. We illustrate the benefits of our proposed training mechanism, leading to considerable improvements in performance in the few-shot setting. Given a single annotated patient, an average Dice score of 0.58 is achieved for the segmentation of ischemic lesions.
BibTeX
@inproceedings{tomasetti2023self,
  author = {Luca Tomasetti and Stine Hansen and Mahdieh Khanmohammadi and Kjersti Engan and Liv Jorunn Høllesli and Kathinka Dæhli Kurz and Michael Kampffmeyer},
  title = {{Self-Supervised Few-Shot Learning for Ischemic Stroke Lesion Segmentation}},
  booktitle = {ISBI},
  year = {2023},
  url = {https://openreview.net/forum?id=QFhS3qprRF},
}
Figure