Methodology for analyzing the causes of photo image misclassification by computer vision programs based on artificial intelligence in medical imaging

Background. The introduction of artificial intelligence (AI)-driven computer vision programs (CVPs) into medical diagnostics has imposed stricter requirements on the quality of input photographic images. Imaging conditions vary significantly across medical fields, necessitating the establishment of field-specific reference ranges for photometric and textural parameters to ensure that AI models maintain reproducible diagnostic accuracy. A systematic analysis of the causes of erroneous classifications by CVPs is essential for their clinical application and further improvement.

Objective: To develop a generalized methodology for analyzing the causes of classification errors in photographic images by processed by AI-driven CVPs. The proposed methodology enables the establishment of reference ranges for photometric and textural parameters for specific medical imaging applications, as well as the development of criteria for excluding images with anomalous values when during preprocessing in medical software systems.

Material and methods. The methodology includes eight sequential stages. A dataset of photographic images verified by histological and dermatoscopic examinations was compiled, classified using the CVPs, and assigned to either of the four standard categories (true positive, true negative, false positive, and false negative). For each photographic image, thirteen photometric and textural quality metrics were calculated, including: brightness, contrast, sharpness, entropy, high-frequency saturation, proportions of overexposed and underexposed pixels, and mean values ​​and standard deviations of the color channels. Systematic between-group differences were identified using one-way ANalysis Of VAriance (ANOVA), Welch's test, and Spearman's rank correlation analysis. The image regions that determine the neural network decision were localized using explainable AI techniques. Reference ranges were established from the characteristics of correctly classified photographs (true positive and true negative categories), defined as intervals of [mean − 2 std; mean + 2 std]. The effectiveness of parameter normalization was assessed by the improvement in accuracy, sensitivity, and specificity.

Results. The proposed methodology was tested using the Derma Onko Check and Melanoma Check CVPs as an example. Its application allowed statistically significant intergroup differences in photometric and textural parameters to be identified (F=13.50–39.31, p<0.001 for the main metrics of the one-way ANOVA; F=5.41–72.29, p<0.001 for the conclusion category by the two-way ANOVA). The analysis confirmed that the observed patterns were independent of the specific CVP used (p=0.39–0.96 for the program factor; p=0.15–0.92 for the interaction effect). Multivariate analysis further demonstrated significant differences among classification outcome groups based on the combined set of image quality metrics (Wilks' lambda 0.639; F=10.37; p<0.001) and established key independent predictors of classification errors through logistic regression (Fast Fourier Transform blur: odds ratio (OR)  3.08; sharpness: OR 0.31; proportion of overexposed pixels: OR 1.64). Reference ranges were established for brightness (0.467–0.942), contrast (0.066–0.333), entropy (3.626–5.590), and high-frequency saturation (23.82–56.48), along with critical thresholds for image exclusion from inference (proportion of overexposed or darkened pixels greater than 55%). The use of a targeted preprocessing module for normalization of deviating parameters falling outside the reference ranges ensured an increase in diagnostic accuracy by +0.014–0.017 in absolute values across all studied CVP configurations, with a predominant increase in specificity (+0.015–0.019).

Conclusion. The proposed methodology for analyzing the causes of erroneous classification of photographic images by AI-driven CVPs was tested on the example of Derma Onko Check and Melanoma Check using a limited dataset (460 photographic images representing two morphological subgroups of melanocytic skin tumors). The extension of the methodology to other areas of medical imaging (ophthalmology, histology, or ultrasound diagnostics), where image acquisition conditions differ substantially, will require validation on representative multicenter datasets with recalculation of parameter reference ranges to reflect the imaging specifics of each domain. The integration of modules for detecting and excluding images with abnormal metric values constitutes a natural practical implication of the proposed methodology and ensures a reproducible increase in the clinical accuracy of AI-driven CVP systems.

1. Lamotkin A.I., Korabelnikov D.I., Olisova O.Yu., Lamotkin I.A. Effectiveness of preliminary differential diagnosis of benign and malignant skin neoplasms using the Derma Onko Check artificial intelligence program. FARMAKOEKONOMIKA. Sovremennaya farmakoekonomika i farmakoepidemiologiya / FARMAKOEKONOMIKA. Modern Pharmacoeconomics and Pharmacoepidemiology. 2025; 18 (2): 261–70 (in Russ.). https://doi.org/10.17749/2070-4909/farmakoekonomika.2025.294.

2. Nadour N., Duguet T., Zahedi S., et al. Diagnostic accuracy of artificial intelligence compared to family physicians and dermatologists for skin conditions: a systematic review and meta-analysis. BMC Prim Care. 2025; 26 (1): 384. https://doi.org/10.1186/s12875-025-03073-9.

3. Zapata M.A., Royo-Fibla D., Font O., et al. Artificial intelligence to identify retinal fundus images, quality validation, laterality evaluation, macular degeneration, and suspected glaucoma. Clin Ophthalmol. 2020; 14: 419–29. https://doi.org/10.2147/OPTH.S235751.

4. Prasad P., Khair A.M.B., Saeed M., Shetty N. Artificial intelligence in histopathology. J Pharm Bioallied Sci. 2024; 16 (Suppl. 5): S4226–9. https://doi.org/10.4103/jpbs.jpbs_727_24.

5. Khalaf K., Terrin M., Jovani M., et al. A comprehensive guide to artificial intelligence in endoscopic ultrasound. J Clin Med. 2023; 12 (11): 3757. https://doi.org/10.3390/jcm12113757.

6. Qamar T., Bawany N.Z. Understanding the black-box: towards interpretable and reliable deep learning models. Peer J Comput Sci. 2023; 9: e1629. https://doi.org/10.7717/peerj-cs.1629.

7. Evans H., Snead D. Understanding the errors made by artificial intelligence algorithms in histopathology in terms of patient impact. NPJ Digit Med. 2024; 7 (1): 89. https://doi.org/10.1038/s41746-024-01093-w.

8. Korabelnikov D.I., Lamotkin A.I. Analysis of the reasons of misclassification of non-melanocytic skin tumors by artificial intelligence based programs. Siberian Journal of Clinical and Experimental Medicine. 2026; 41 (1): 221–31 (in Russ.). https://doi.org/10.29001/2073-8552-2026-2881.

9. Daneshjou R., Vodrahalli K., Novoa R.A., et al. Disparities in dermatology AI performance on a diverse, curated clinical image set. Sci Adv. 2022; 8 (32): eabq6147. https://doi.org/10.1126/sciadv.abq6147.

10. Daneshjou R., Smith M.P., Sun M.D., et al. Lack of transparency and potential bias in artificial intelligence data sets and algorithms: a scoping review. JAMA Dermatol. 2021; 157 (11): 1362–9. https://doi.org/10.1001/jamadermatol.2021.3129.

11. Liopyris K., Gregoriou S., Dias J., Stratigos A.J. Artificial Intelligence in dermatology: challenges and perspectives. Dermatol Ther. 2022; 12 (12): 2637–51. https://doi.org/10.1007/s13555-022-00833-8.

12. Haggenmüller S., Maron R.C., Hekler A., et al. Skin cancer classification via convolutional neural networks: systematic review of studies involving human experts. Eur J Cancer. 2021; 156: 202–16. https://doi.org/10.1016/j.ejca.2021.06.049.