Artificial Intelligence and Chest X-ray Analysis: A Study Conducted by Researchers from YULCOM Technologies, Concordia University, and the McGill University on Multi-Label Classification of Pulmonary Diseases
The R&D team at YULCOM Technologies contributed to the research and authorship of the paper titled “Clinically-Inspired Hierarchical Multi-Label Classification of Chest X-rays with a Penalty-Based Loss Function.”
This work, co-authored by YULCOM’s R&D team in collaboration with Concordia University (Gina Cody School of Engineering and Computer Science) and the Division of Radiation Oncology at the McGill University Health Centre, was supported by a Canadian NSERC Discovery Grant (RGPIN-2017-06722).
The paper is available on arXiv.
Researchers Mehrdad Asadi, Komi Sodoké, PhD, Ian J. Gerard, and Marta Kersten-Oertel, PhD, propose an artificial intelligence (AI) approach for multi-label classification of chest X-rays (CXR), aiming to improve diagnostic accuracy while enhancing clinical interpretability.
In the medical field, AI is increasingly positioned as a decision-support tool to address key challenges such as radiologist workload, variability in interpretations, and the risk of diagnostic errors. However, conventional AI approaches face important limitations, particularly their inability to model clinical relationships between diseases and their lack of transparency.
To overcome these challenges, the research team introduces a clinically inspired hierarchical approach that mirrors medical reasoning. Diseases are structured according to a hierarchy reflecting real-world pathological relationships. In addition, the authors propose a novel loss function—Hierarchical Binary Cross-Entropy (HBCE)—which penalizes inconsistencies in predictions (for example, predicting a specific condition without its broader category), thereby producing results that are more clinically coherent.
The proposed model is based on a deep learning architecture trained on the CheXpert dataset. It also incorporates mechanisms to enhance interpretability, including:
- visualization of relevant regions using Grad-CAM,
- uncertainty estimation through Monte Carlo dropout.
The results demonstrate strong performance (AUROC ≈ 0.903), with notable improvements when data-driven penalty mechanisms are applied. Most importantly, the approach generates predictions that are more consistent with clinical reasoning.
This work highlights the value of integrating clinical knowledge into AI models to develop systems that are:
- more reliable,
- more interpretable,
- and better suited for real-world clinical use.
Ultimately, AI is presented not as a replacement for physicians, but as a complementary tool capable of improving both the quality and efficiency of medical diagnosis.
