¹Department of Neuropathology, Pathology, Medical Faculty, University of Augsburg, Augsburg, Germany
²DigiHealth Institute, Neu-Ulm University of Applied Sciences, Neu-Ulm, Germany
³Institute of Medical Data Science, University Hospital of Würzburg, Würzburg, Germany
⁴Department of Dermatology, Venereology and Allergology, University Hospital Würzburg, Würzburg, Germany
⁵Chair of Computer Science II - Software Engineering, University of Würzburg, Würzburg, Germany
⁶Faculty of Applied Computer Science, University of Augsburg, Augsburg, Germany
⁷Institute for Digital Medicine, University Hospital Augsburg, Augsburg, Germany

Text

Introduction: While using Computer Vision (CV) to classify images of wounds is no new task [1], [2], rare diseases are often omitted due to scarce data. However, particularly in medical diagnostics, the most diagnostically relevant images are often those that are more difficult to classify due to their complexity or subtle pathological features. While previous work focused on differentiation of single rare wound types [3], we present a CV classification model differentiating between images of the rare wound diseases Livedo Vasculopathy (LV), Pyoderma Gangraenosum (PG), and Ulcus Hypertonicum Martorell (UHM) as well as the common Venous Leg Ulcer (VLU) wounds.

Methods: A multicentric dataset was scanned by trained dermatologists from the University Hospital of Würzburg. Bias-introducing artifact (i.e. tattoos, measuring tapes) were removed and images were cropped to contain relevant wound areas and healthy tissue with the least amount of background (i.e. parts of the wall, op-sheet). All cases were confirmed both clinically as well as pathologically. The final dataset contained 2,921 wound-images from 104 patients (LV 204 images, 17 patients; PG 1,304 images, 49 patients; UHM 305 images, 28 patients; and VLU 1,108 images, 10 patients).

Using the AUCMEDI framework [4], a ConvNeXt-based [5] multi-class classification pipeline was trained for separating the four wound classes. A class-weighted categorical focal loss and a patient stratified 80/20% split were used for training and testing. In total, three models were trained using 3-fold cross-validation. Testing was conducted using an ensemble of three models, with predictions obtained by averaging. Code and results: https://github.com/hnu-digihealth/GMDS_Wound_Type_Classification

Results: The resulting ensemble expresses a high class-wise AUC (LV: 81.50%, PG: 85.44%, UHM: 86.07, ULV: 99.29%). Regarding the sensitivity and specificity of the rare wound diseases, LV and UHM expressed a high specify (LV: 99.60%, UHM: 88.38%) with a low sensitivity (LV: 36.17%, UHM: 45.61%). Both were often confused with PG (LV 38.30/, UHM 52.63%), resulting in reduced results for PG (specificity: 76.08%, sensitivity: 73.03%). The differentiation ability between rare diseases (LV, UHM, PG) and common wounds (VUL) is shown by VUL’s specificity of 95.34% and sensitivity of 97.14%.

While the results are promising for a first classification study, the sensitivity for the rare diseases is a major concern. Taking a detailed look at the data (LV 204 images, 17 patients; UHM 305 images, 28 patients; and PG 1,304 images, 49 patients) the number of images and patients for both LV and UHM was < 50% of the data available for PG. This could explain the model’s preference to select PG for rare wound diseases. As the collection of a larger dataset is difficult for rare diseases, optimization in the preprocessing and augmentation will be conducted to further improve the differentiation performance.

Conclusion: This work shows a first proof of concept for an ML-model-based differentiation between 3 rare wound diseases and a common one. While the differentiation between rare and common is of excellent quality, the internal differentiation of the three rare wound types is still improvable. Further research should focus on increasing the model’s sensitivity for the rare diseases.

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.

References

[1] Anisuzzaman DM, Patel Y, Rostami B, Niezgoda J, Gopalakrishnan S, Yu Z. Multi-modal wound classification using wound image and location by deep neural network. Sci Rep. 2022 Nov 21;12(1):20057. DOI: 10.1038/s41598-022-21813-0
[2] Secco J, Spinazzola E, Pittarello M, Ricci E, Pareschi F. Clinically validated classification of chronic wounds method with memristor-based cellular neural network. Sci Rep. 2024 Dec 28;14(1):30839. DOI: 10.1038/s41598-024-81521-9
[3] Birkner M, Schalk J, Von Den Driesch P, Schultz ES. Computer-Assisted Differential Diagnosis of Pyoderma Gangrenosum and Venous Ulcers with Deep Neural Networks. JCM. 2022 Nov 30;11(23):7103. DOI: 10.3390/jcm11237103
[4] Müller D, Hartmann D, Soto-Rey I, Kramer F. Abstract: AUCMEDI. In: Bildverarbeitung für die Medizin 2023. Wiesbaden: Springer Vieweg; 2023 [cited 2025 Apr 14]. p. 253–253. DOI: 10.1007/978-3-658-41657-7_55
[5] Liu Z, Mao H, Wu CY, Feichtenhofer C, Darrell T, Xie S. A ConvNet for the 2020s. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2022. 2022 [cited 2023 Sep 8]. p. 11976–86. Available from: https://openaccess.thecvf.com/content/CVPR2022/html/Liu_A_ConvNet_for_the_2020s_CVPR_2022_paper.html