The article Digital Image Processing and Convolutional Neural Network Applied to Detect Mitral Stenosis in Echocardiograms: Clinical Decision Support by Barros Filho et al. presents an investigation into the application of computational imaging methods for assisting clinicians in the diagnosis of mitral stenosis. The authors propose a combined approach using digital image processing (DIP) techniques alongside convolutional neural networks (CNNs), aiming to provide a supportive framework for interpreting transoesophageal echocardiography. Their work demonstrates methodological promise while also highlighting the current challenges of applying artificial intelligence in clinical cardiology.
Keywords: Mitral stenosis, echocardiography, convolutional neural networks, digital image processing, clinical decision support, artificial intelligence
Introduction
Cardiovascular disease remains the leading cause of death worldwide, with valvular disorders such as mitral stenosis contributing significantly to global morbidity. Echocardiography is the primary imaging tool for diagnosing these conditions, yet its interpretation requires a high level of expertise and can be influenced by operator experience. In parallel, the rapid progress of artificial intelligence has opened new opportunities for medical imaging, where algorithms can detect patterns that may be overlooked or inconsistently recognised by human observers. Techniques such as digital image processing and convolutional neural networks are increasingly being applied to cardiac imaging, with the aim of providing clinicians with decision-support systems that improve accuracy, reduce variability, and ultimately enhance patient care.
Strengths
The study addresses a clinically significant problem: reliable and automated identification of mitral stenosis, a condition where accurate interpretation of echocardiograms often depends on specialist expertise. By combining DIP with CNNs, the work demonstrates the potential of computational methods in clinical decision support. The reported CNN accuracy of 92% for distinguishing between cases with and without stenosis is promising. Moreover, the inclusion of DIP to calculate valve opening areas adds interpretability, providing clinicians with quantitative measures that can aid validation.
The authors also situate their work within the broader literature, contrasting their approach with studies based on phonocardiograms or electrocardiograms. Their choice of echocardiographic video analysis makes the study directly relevant to imaging specialists and cardiologists.
Limitations
A significant limitation lies in the dataset size. With only 30 echocardiogram videos, the training set is too limited to robustly develop a CNN capable of fine-grained classification across mild, moderate, and severe stenosis. While data augmentation partially mitigates this, synthetic transformations cannot substitute for the diversity and variability of real-world clinical data. This restriction is evident in the model’s inability to reliably discriminate between different stenosis severities, reducing its utility in nuanced diagnostic scenarios.
The reliance on a single, relatively simple segmentation method (single-band fixed thresholding) also raises concerns. While effective for binary separation, the method is prone to variability depending on imaging conditions and device-specific parameters. Alternative segmentation techniques, such as adaptive thresholding or deep-learning–based segmentation, could provide more robust feature extraction and reduce dependence on manual calibration.
Another limitation is the interpretability of the CNN output. While DIP provides contour delineation and quantitative metrics, the CNN only delivers binary classification. For clinical adoption, models must provide explainable outputs that clinicians can scrutinise alongside imaging data, especially in borderline or ambiguous cases.
Clinical and Research Implications
The findings show that DIP alone can achieve multi-class classification, while the CNN offers strong binary discrimination. A hybrid model, integrating DIP-derived metrics with CNN features, may offer a path forward. Importantly, expanding the dataset through multicentre collaboration is essential to validate generalisability, especially since mitral stenosis prevalence and imaging quality vary across populations and equipment.
For future studies, emphasis should be placed on:
- Acquiring a substantially larger dataset with balanced classes.
- Exploring more sophisticated segmentation methods beyond fixed thresholds.
- Enhancing interpretability of CNN predictions to build clinician trust.
- Evaluating performance across different imaging systems (Philips, GE, etc.) to ensure real-world applicability.
Conclusion
This article makes a valuable contribution to the growing field of AI-assisted echocardiography by demonstrating the feasibility of combining DIP and CNNs for detecting mitral stenosis. However, the limited dataset and methodological simplifications constrain its clinical applicability at this stage. With further refinement and larger-scale validation, the proposed framework could become a valuable decision-support tool, reducing diagnostic variability and supporting cardiologists in complex assessments.
Reference
Barros Filho, G. de F.; Firmino, J.F. de M.; Solha, I.; Medeiros, E.F. de; Felix, A. de S.; Lima Júnior, J.C. de; Melo, M.D.T. de; Rodrigues, M.C. Digital Image Processing and Convolutional Neural Network Applied to Detect Mitral Stenosis in Echocardiograms: Clinical Decision Support. Journal of Imaging 2025, 11(8), 272. https://doi.org/10.3390/jimaging11080272
Disclaimer
This article provides a critical evaluation of the study Digital Image Processing and Convolutional Neural Network Applied to Detect Mitral Stenosis in Echocardiograms: Clinical Decision Support by Barros Filho et al. It is intended for academic and informational purposes only and does not constitute medical advice, clinical guidance, or a substitute for professional healthcare consultation. The research discussed remains subject to the limitations of the original study, including dataset size and methodological constraints. Readers are advised that the approaches described are investigational and should not be applied in clinical practice without appropriate validation, peer review, and regulatory approval. The author(s) of this article accept no liability for any interpretation or use of the content presented herein.