Peer-reviewed articles
[1] A. E. Sultana, M. Oniga, P. F. Rus, A. L. Dobre and O. A. Orzan,
"MelanoDet: Multimodal Imaging System for Screening and Analysis of Cutaneous Melanoma,"
in IEEE Access, vol. 13, pp. 61469–61482, 2025, doi: 10.1109/ACCESS.2025.3557628.
This article presents the overall MelanoDet concept and implementation of the VIS/NIR/thermal imaging system for cutaneous lesions. It describes the hardware design, the clinical acquisition workflow, and the structure of the resulting multimodal dataset, illustrated with example cases and preliminary analyses on the collected images.
This article presents the overall MelanoDet concept and implementation of the VIS/NIR/thermal imaging system for cutaneous lesions. It describes the hardware design, the clinical acquisition workflow, and the structure of the resulting multimodal dataset, illustrated with example cases and preliminary analyses on the collected images.
[2] M. Oniga, M. Mihăilescu, A. E. Sultana, M. Chivu, I. Tudose, O. A. Orzan, and A. Marin,
"Hyperspectral imaging-based framework for automated detection of dysplastic melanocytes in dermatopathology,"
in 18th International Conference on Machine Vision (ICMV2025), in press, 2025.
This work proposes a machine-learning pipeline for hyperspectral pathology slides to automatically detect melanocytes and identify NIR wavelengths that can distinguish between nevus, dysplastic nevus, and melanoma. Using unsupervised clustering and post-processing techniques (Fuzzy C-means and Mean Shift), the authors extract NIR spectra and train several classifiers, achieving best accuracies around 0.74. Mutual information identifies five highly discriminative NIR wavelengths for future multispectral systems.
This work proposes a machine-learning pipeline for hyperspectral pathology slides to automatically detect melanocytes and identify NIR wavelengths that can distinguish between nevus, dysplastic nevus, and melanoma. Using unsupervised clustering and post-processing techniques (Fuzzy C-means and Mean Shift), the authors extract NIR spectra and train several classifiers, achieving best accuracies around 0.74. Mutual information identifies five highly discriminative NIR wavelengths for future multispectral systems.
[3] M. Oniga, P. F. Rus, R. Condorovici, A. Marin, C. Cotrut, D. O. Costache, H. Blejan,
R. C. Jecan, C. Corciu, and A. E. Sultana,
"Heatmap-based deep learning framework for multi-modal registration of VIS, NIR, and thermal images in dermatological imaging,"
Digital Signal Processing, under review, 2025.
This paper focuses on the image-registration pipeline that turns raw VIS, NIR and thermal images into well-aligned triplets. It details the geometry, the homography-heatmap alignment strategy, and similarity metrics such as SSIM, MI, and NMI to assess registration quality.
This paper focuses on the image-registration pipeline that turns raw VIS, NIR and thermal images into well-aligned triplets. It details the geometry, the homography-heatmap alignment strategy, and similarity metrics such as SSIM, MI, and NMI to assess registration quality.
Patent Application
[1] Sistem inteligent de evaluare și interpretare automată a nevilor displazici bazat pe imagistică multispectrală.
This patent application covers a portable multispectral system combining synchronized VIS, NIR and thermal imaging with AI-based analysis for automated assessment of pigmented skin lesions. It uses inter-sensor calibration, registration, and ML-driven fusion to output interpretable malignancy risk maps.
This patent application covers a portable multispectral system combining synchronized VIS, NIR and thermal imaging with AI-based analysis for automated assessment of pigmented skin lesions. It uses inter-sensor calibration, registration, and ML-driven fusion to output interpretable malignancy risk maps.