Editorial: Artificial intelligence-based medical image automatic diagnosis and prognosis prediction

Junchi Yan, Yu‐Kun Lai, Yi Xu, Yinqiang Zheng, Zhibin Niu, Tao TanFrontiers in Physics 2023 · DOI · OpenAlex

Introduction

This editorial provides a high-level synthesis of a specialized "Research Topic" collection published in Frontiers in Physics. It highlights nine peer-reviewed manuscripts that advance the application of artificial intelligence (AI)—specifically deep learning and radiomics—to medical imaging for both automated diagnosis and long-term prognosis. By tackling fundamental challenges in the field, these studies demonstrate significant performance gains across various clinical domains, from neuroimaging to oncology.

Addressing Critical Clinical Gaps

While deep learning has revolutionized medical imaging, several "last-mile" challenges prevent these tools from reaching their full clinical potential. The research featured in this collection identifies and addresses four primary gaps:

  1. Limited Annotations: Overcoming the difficulty and high cost of obtaining high-quality, expert-labeled medical data.
  2. Small Object Detection: Enhancing the sensitivity of standard algorithms that often miss tiny but critical features, such as early-stage lesions or specific cell types.
  3. Complex Data Integration: Developing methods to combine information from different modalities, such as pairing PET scans with MRI or clinical metadata.
  4. Prognosis vs. Diagnosis: Moving beyond simple disease identification to predicting future disease trajectories and patient outcomes.

Methodological Innovations

The editorial categorizes the featured research into several cutting-edge methodological frameworks designed to improve diagnostic precision:

  • Advanced Architectures: The use of ResAttn-Recon (a residual self-attention encoder-decoder) for brain MR reconstruction and FunSwin (a Swin Transformer-based model) for fundus image analysis.
  • Hybrid Learning: The LAD-GCN framework combines Graph Convolutional Networks (GCN) to capture spatial relationships between cells with Convolutional Neural Networks (CNN) for global semantic features in histopathology.
  • Loss Function Innovation: The introduction of Pixel Level Balancing (PLB) weights training loss based on the number of pixels in a detection box, specifically to improve the detection of small objects that are often "washed out" in standard training.
  • Radiomics & Machine Learning: Traditional feature engineering (radiomics) is paired with modern classifiers like XGBoost and the Boruta algorithm for robust feature selection in CT and PET imaging.

Key Insights and Quantitative Results

The research demonstrates that architectural innovation leads to measurable improvements in clinical tasks. Key milestones highlighted include:

  • Neuroimaging: The ResAttn-Recon model achieved an average Hausdorff distance of 0.629 for cortical surface reconstruction, significantly outperforming the previous state-of-the-art model, DeepCSR (0.746).
  • Oncology (Liver): A hybrid radiomics model for predicting microvascular invasion achieved a near-perfect AUC of 0.996 by combining PET/CT data with contrast-enhanced MRI, suggesting that multi-modal models are the future of oncology.
  • Oncology (Kidney): An XGBoost model for predicting lymph node metastasis reached an AUC of 0.916, identifying pathological grade and T-stage as the most critical predictive variables.
  • Graph-Based Analysis: In lung cancer studies, treating cells as nodes in a graph (GCN) rather than pixels in a grid allowed models to better understand the "architecture" of a tumor, leading to superior growth pattern estimation.

Future Directions

While the results are promising, the editorial identifies several hurdles for the medical AI community to overcome:

  • Generalization: Many studies rely on specific clinical cohorts; validating these models across more diverse, multi-center populations is a necessary next step.
  • Annotation Bottlenecks: The field continues to struggle with the high cost of expert data labeling, necessitating further research into semi-supervised and weakly-supervised techniques.
  • Clinical Integration: The transition from high AUC scores in a research setting to functional, bedside tools remains a significant challenge for real-world implementation.

Summary updated

Abstract

EDITORIAL article Front. Phys., 25 May 2023Sec. Medical Physics and Imaging Volume 11 - 2023 | https://doi.org/10.3389/fphy.2023.1210010
BibTeX
@inproceedings{yan2023editorial,
  author = {Junchi Yan and Yu‐Kun Lai and Yi Xu and Yinqiang Zheng and Zhibin Niu and Tao Tan},
  title = {{Editorial: Artificial intelligence-based medical image automatic diagnosis and prognosis prediction}},
  booktitle = {Frontiers in Physics},
  year = {2023},
  doi = {10.3389/fphy.2023.1210010},
}

Loading related content...

Figure