Hybrid Learning System for Large-scale Medical Image Analysis

Zehua Cheng, Lianlong WuIJCAI 2022 · DOI

TL;DR

Medical imaging analysis is hampered by the high cost of annotation. This paper proposes a "hybrid learning system" that combines self-supervised, semi-supervised, and weak-supervised learning to reduce reliance on high-quality annotations. The goal is to enable large-scale medical image analysis even with limited or imperfectly labeled data.

What Problem Does This Paper Address?

Deep learning models, especially in medical imaging, thrive on vast amounts of high-quality annotated data. However, obtaining such data is a significant bottleneck due to the need for expert knowledge, time, and cost. This scarcity limits the application of deep learning to large-scale medical image analysis tasks. The authors aim to bridge this gap by developing methods that can effectively learn from medical images with varying degrees of annotation quality, from precise segmentation masks to coarser labels or even just bounding boxes that might not perfectly outline the target.

Methods

The core of the proposed solution is a hybrid learning system that integrates three learning paradigms:

  • Self-supervised learning: This approach trains a model on pretext tasks (e.g., reconstruction) using unlabeled data to learn generalizable features. This is a standard technique for leveraging large unlabeled datasets.
  • Semi-supervised learning: This paradigm learns from a small amount of labeled data alongside a large amount of unlabeled data. It acts as a bridge between fully supervised and self-supervised learning.
  • Weak-supervised learning: This method trains models using less precise annotations, such as bounding boxes instead of pixel-level segmentation masks.

The authors also explore graph constraints to further enhance the hybrid system. They suggest that incorporating knowledge graphs can help organize information more effectively and improve annotation utilization by modeling domain knowledge and expert rules. They also mention investigating symbolic rule learners for inductive relational learning.

Datasets

The paper mentions using the SIIM-FISABIO-RSNA COVID-19 Detection benchmark for evaluating their methods on medical image detection tasks under different annotation qualities. Specific details about other datasets used for training or evaluation of the hybrid system are not extensively detailed, but the general context implies the use of large-scale medical image datasets with varying annotation types.

Key Insights and Ideas

  • Annotation quality significantly impacts model performance: The paper demonstrates that even slight inaccuracies in bounding box annotations lead to a noticeable degradation in detection performance, as shown in Table 1.
  • Hybrid learning is a promising direction: Combining self-supervised, semi-supervised, and weak-supervised learning offers a robust way to tackle annotation scarcity.
  • Semi-supervised learning as a bridge: Preliminary experiments suggest that semi-supervised learning can effectively connect self-supervised and weak-supervised approaches, improving overall performance and mitigating localization sacrifices.
  • Knowledge graphs can enhance medical image analysis: Integrating knowledge graphs into the hybrid system is proposed as a way to leverage domain knowledge and expert rules, potentially improving annotation utilization and model performance.

Key Results

The paper presents experimental results in Table 1, showing the average precision (%) of different detection models (Faster R-CNN, Cascade R-CNN, RetinaNet) on the SIIM-FISABIO-RSNA COVID-19 Detection benchmark. When moving from "Baseline" (normal annotation) to "Inaccurate Annotation," the average precision drops significantly for all models:

  • Faster R-CNN: 67.3% to 54.3%
  • Cascade R-CNN: 66.4% to 55.4%
  • RetinaNet: 61.3% to 53.2%

This clearly illustrates the problem of performance degradation with imperfect annotations. The paper states that their developed regularization policy and adversarial loss can improve robustness and performance under such conditions, though specific quantitative results for these improvements are not detailed in the provided text.

Limitations and Future Work

Limitations: * The paper acknowledges that manual annotations are not always exact due to practical constraints, leading to inexact bounding boxes. * While they present preliminary results, a comprehensive evaluation of the full hybrid system across various medical imaging tasks and annotation types is implied to be ongoing.

Future Work: * The authors plan to further explore the potential of their hybrid learning system over large-scale data. * They intend to collaborate with medical institutions to further verify their design and ensure its practical applicability. * They aim to integrate graph constraints and symbolic rule learners to enhance the hybrid system's precision and efficiency.

Summary updated

BibTeX
@inproceedings{cheng2022hybrid,
  author = {Zehua Cheng and Lianlong Wu},
  title = {{Hybrid Learning System for Large-scale Medical Image Analysis}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.24963/ijcai.2022/824},
}
Figure