Embracing Imperfect Datasets: A Review of Deep Learning Solutions for Medical Image Segmentation (1908.10454v2)

Abstract: The medical imaging literature has witnessed remarkable progress in high-performing segmentation models based on convolutional neural networks. Despite the new performance highs, the recent advanced segmentation models still require large, representative, and high quality annotated datasets. However, rarely do we have a perfect training dataset, particularly in the field of medical imaging, where data and annotations are both expensive to acquire. Recently, a large body of research has studied the problem of medical image segmentation with imperfect datasets, tackling two major dataset limitations: scarce annotations where only limited annotated data is available for training, and weak annotations where the training data has only sparse annotations, noisy annotations, or image-level annotations. In this article, we provide a detailed review of the solutions above, summarizing both the technical novelties and empirical results. We further compare the benefits and requirements of the surveyed methodologies and provide our recommended solutions. We hope this survey article increases the community awareness of the techniques that are available to handle imperfect medical image segmentation datasets.

• The paper presents a comprehensive review of strategies that mitigate scarce and weak annotations in medical image segmentation.
• It examines methods like data augmentation, transfer learning, and self-supervised approaches to enhance segmentation performance.
• The review highlights potential hybrid techniques that could accelerate clinical deployment of AI-driven medical imaging.

Overview of Deep Learning Solutions for Medical Image Segmentation with Imperfect Datasets

The paper "Embracing Imperfect Datasets: A Review of Deep Learning Solutions for Medical Image Segmentation" presents a comprehensive review of methodologies addressing the challenge of training deep learning models using imperfect datasets. Specifically, it focuses on two major issues: scarce annotations and weak annotations in medical imaging.

Challenges in Medical Image Segmentation

Deep learning models, particularly convolutional neural networks (CNNs), have set performance benchmarks in medical image segmentation tasks. However, these models typically require extensive, high-quality annotated datasets, which are rarely available in medical imaging due to the high cost and complexity of generating such annotations. Consequently, researchers have explored various strategies to mitigate the limitations posed by imperfect datasets.

Scarce Annotations

Scarce annotations refer to datasets with limited annotated examples. The paper identifies several strategies to address this issue:

1. Data Augmentation: Traditional techniques such as spatial and intensity transformations have been used extensively to artificially expand datasets. More sophisticated methods involve mixing images or generating synthetic data using GANs, creating additional training examples without significant correlation to the original data.
2. Transfer Learning and Domain Adaptation: These methods leverage pre-trained models or labeled datasets from related domains. Transfer learning allows for the reuse of networks trained on large natural image datasets (e.g., ImageNet). Domain adaptation involves aligning feature distributions between source and target domains.
3. Semi-Supervised and Self-Supervised Learning: These approaches make use of large amounts of unlabeled data alongside limited labeled data. Self-supervised methods generate surrogate tasks to learn meaningful representations, which are then fine-tuned for segmentation. Semi-supervised learning may utilize pseudo labels generated iteratively to improve the model.
4. Regularized Training: Techniques such as multi-task learning and shape regularization ensure that models generalize better by imposing helpful constraints during training.
5. Post-Segmentation Refinement Using CRFs: Variants of Conditional Random Fields are employed to refine segmentation outputs, though their effectiveness in 3D segmentation tasks has shown mixed results.

Weak Annotations

Weak annotations cover scenarios where annotations are sparse or noisy:

1. Image-Level Labels: Class activation maps (CAMs) and multiple instance learning (MIL) are used to extrapolate pixel-level segmentation from image-level labels, significantly reducing annotation efforts.
2. Sparse Annotations: Techniques involve selective loss functions, applying loss only to labeled pixels, sometimes combined with methods to estimate and refine partially labeled regions iteratively.
3. Noisy Annotations: Robust loss functions are proposed to minimize the impact of label noise, often utilizing learning strategies that identify and downplay unreliable annotations.

Implications and Future Directions

The techniques reviewed offer practical approaches to mitigate the lack of data, a significant bottleneck in deploying medical AI solutions. Moving forward, the integration of these methods could lead to AI systems that require fewer annotations while maintaining high performance, thus accelerating clinical deployment.

The paper's thorough examination of current methodologies guides researchers in selecting suitable approaches based on specific annotation limitations and available resources. Furthermore, the exploration of hybrid approaches combining strategies from different categories could unlock further advancements in leveraging imperfect datasets for medical image segmentation.