Analysis of the reasons of misclassification of non-melanocytic skin tumors by artificial intelligence based programs

Introduction. Differential diagnosis of non-melanocytic skin tumors remains a key challenge in dermato-oncology, as timely detection of malignant forms increases the chances of successful treatment. The subjectivity of traditional methods encourages the use of artificial intelligence (AI), but errors in computer vision programs require analysis.

Aim: To analyze the reasons of misclassification of non-melanocytic skin tumor images by AI-based programs. This was accomplished by identifying systematic differences in data characteristics and visualizing regions of interest during image recognition. The results are aimed at improving the efficiency of training and using computer vision programs.

Material and Methods. Datasets processed in the Derma Onko Check and Melanoma Check programs were used for a retrospective analysis of tumor images. For this study, malignant tumors were considered positive, while benign tumors were considered negative. Considering two types of AI decisions (true and false), four standard result classes were considered: true positive (TP), false positive (FP), true negative (TN), and false negative (FN). To visualize image quality metrics (brightness, contrast, entropy, blur, and RGB metrics), boxplots, paired scatterplots, and pixel difference maps were used. To visualize areas that significantly influence image classification, two explainable AI methods were applied: weighted class activation mapping (Score-CAM) and occlusion sensitivity. These methods allow us to understand which image regions are most important for the classification decisions of the deep neural network. To test statistical hypotheses, Welch's t-test and one-way analysis of variance were used; to assess the relationship between characteristics, Spearman's correlation analysis was used.

Results. Significant differences in image characteristics were identified. The IP results were characterized by the following features. Brightness was lower (median 0.6914 on a normalized scale of 0-1), indicating natural, uniform illumination without strong glare. Entropy was high (median 4.8584), indicating a complex texture with many clinically significant details: ulceration, irregular borders, and pigmentation variations. Blurring was moderate, providing acceptable image sharpness without severe blurring of the tumor edges and texture. The mean values of the red and green channels were balanced. The LP results had increased brightness (median 0.7994, indicating an overexposed, overly bright photo with glare, where fine textural details are lost) and low entropy (median 4.6414, indicating a uniform texture without complex patterns). Significant differences between classes were confirmed for brightness (F = 5.1848; p < 0.05), entropy (F = 5.2509; p < 0.05), FFT blur (F = 3.1136; p < 0.05), green channel mean (F = 5.3315; p < 0.05), and red channel mean (F = 3.3812; p < 0.05). The Score-CAM and Occlusion Sensitivity explainable AI methods and image quality analysis showed that non-melanocytic tumor classification errors by the Derma Onko Check and Melanoma Check AI programs occurred due to overexposure, low entropy, and photography artifacts; false positives occurred on bright, low-texture images, and false negatives occurred on dark/blurred images. AI models are distracted by the background, hair, shadows.

Conclusion. When training computer vision programs, developers are advised to perform image preprocessing (automatic white balance, gamma correction, Sobel filters for texture enhancement and Wiener filters for blur suppression, online brightness and contrast augmentation), normalize color channels, monitor key quality metrics after each training epoch, and use augmentation that compensates for the negative brightness-entropy correlation and illumination variability. Users of the programs are advised to adhere to standard shooting conditions: uniform diffuse lighting without shadows and glare, luminance < 0.75 on the normalized scale, and the absence of artifacts in the frame; shoot in macro mode from a distance of 8-15 cm, centering the tumor and ensuring entropy > 4.8 and a resolution of 2000-3000 pixels on the longest side; stabilize the camera and activate the automatic white balance function on the shooting device.

1. Stafford H., Buell J., Chiang E., Ramesh U., Migden M., Nagarajan P. et al. Non-melanoma skin cancer detection in the age of advanced technology: a review. Cancers (Basel). 2023;15(12):3094. https://doi.org/10.3390/cancers15123094

2. Ferlay J., Colombet M., Soerjomataram I., Dyba T., Randi G., Bettio M. et al. Cancer incidence and mortality patterns in Europe: Estimates for 40 countries and 25 major cancers in 2018. European Journal of Cancer. 2018;103:356–387. https://doi.org/10.1016/j.ejca.2018.07.005

3. Fahim Y.A., Hasani I.W., Kabba S., Ragab W.M. Artificial intelligence in healthcare and medicine: clinical applications, therapeutic advances, and future perspectives. European Journal of Medical Research. 2025;30(1):848. https://doi.org/10.1186/s40001-025-03196-w

4. Behara K., Bhero E., Agee J.T. AI in dermatology: a comprehensive review into skin cancer detection. PeerJ Computer Science. 2024;10:e2530. https://doi.org/10.7717/peerj-cs.2530

5. Melarkode N., Srinivasan K., Qaisar S.M., Plawiak P. AI-Powered Diagnosis of Skin Cancer: A Contemporary Review, Open Challenges and Future Research Directions. Cancers. 2023;15(4):1183. https://doi.org/10.3390/cancers15041183

6. Kolla L., Parikh R.B. Uses and limitations of artificial intelligence for oncology. Cancer. 2024;130(12):2101–2107. https://doi.org/10.1002/cncr.35307

7. Lamotkin A.I., Korabelnikov D.I., Olisova O.Iu., Lamotkin I.A. Effectiveness of preliminary differential diagnosis of benign and malignant skin neoplasms using the artificial intelligence program Derma Onko Check. Farmakoekonomika. Modern Pharmacoeconomics and Pharmacoepidemiology. 2025;18(2):261–270. (In Russ.). https://doi.org/10.17749/2070-4909/farmakoekonomika.2025.294

8. Lamotkin A.I., Korabelnikov D.I., Lamotkin I.A. et al. Accuracy of preliminary diagnosis of malignant melanocytic skin tumors using an artificial intelligence program. Medical Bulletin of the Main Military Clinical Hospital named after N.N. Burdenko. 2025;(1(19)):42–51. (In Russ.). https://doi.org/10.53652/2782-1730-2025-6-1-42-51

9. Mevorach L., Farcomeni A., Pellacani G., Cantisani C. A comparison of skin lesions’ diagnoses between AI-based image classification, an expert dermatologist, and a non-expert. Diagnostics (Basel). 2025;15(9):1115. –