In this paper, we propose a new method for removing specular reflection from RGB dermatological images to improve skin disease identification. This method is based on the dichromatic reflection model, in which a skin reflectance image is considered as a mixture of diffuse reflection components and a specular reflection component. Based on Blind Source Separation (BSS) and exploiting Artificial Intelligence (AI), the proposed method proceeds in two steps. The first step consists of estimating each of these components separately in a fully blind manner by exploiting BSS methods based on Independent Component Analysis (ICA) and Non-negative Matrix Factorization (NMF), respectively. By retaining only the diffuse reflection image, we obtain a new image, referred to as a pre-processed image that is free of specular reflection. The second step consists of identifying the skin disease from the new pre-processed image using AI. The performance of our method is evaluated on several databases of real RGB dermatological images of melanoma and naevus, as well as on new artificial dermatological images, and then compared with several existing methods. The obtained results show that the proposed method outperforms the existing methods in terms of specular reflection removal in the first pre-processing step. Furthermore, the contribution of this pre-processing step to the identification of the two diseases (melanoma and naevus) was evaluated, and the results show that the exploitation of RGB dermatological images pre-processed by our method significantly improves the identification performance compared to the exploitation of raw RGB dermatological images.
Generative deep learning methods such as Variational Autoencoders, Generative Adversarial Networks or Diffusion Models have achieved remarkable results in image synthesis. In particular, Diffusion Models have demonstrated superior image fidelity and sample diversity compared to other approaches on standard benchmarks, while also offering more flexible training dynamics as evidenced both in natural image domains and in medical applications. In medical imaging including Computed Tomography, Magnetic Resonance Imaging or Positron Emission Tomography, such models have various important use-cases including data augmentation, domain transfer, image reconstruction, inpainting, image augmentation and artifact correction. Further, generated images can be conditioned on patient characteristics, medical reports, segmentation masks or other images. Specific challenges arise due to the 3D volumetric nature of medical images, which translate to substantial demands in terms of memory, compute, and training data. This article provides a comparative overview of 3D Diffusion Models which have been developed for medical images, with a focus on current performance, limitations, and an outlook for future work.
Background and objective Variations in staining protocols for peripheral blood smears lead to inconsistent image coloration, posing significant challenges for automated analysis. This study evaluates the effectiveness of color transfer as a preprocessing step to harmonize blood smear image color spaces and improve white blood cell classification performance using a convolutional neural network (CNN). Method Nine color transfer algorithms were evaluated using a May-Grünwald-Giemsa-stained image as a reference. A two-stage evaluation was employed: first, assessment of the color transfer effectiveness; and second, measurement of the impact of preprocessed images on a CNN model across three backbones (ResNet-50, Inception-V3 and EfficientNet-B3). Results The findings demonstrate that improper color transfer algorithms can inherently degrade clinical feature integrity by introducing unnatural visual elements. Specifically, the color distribution transfer method causes a significant drop in classification accuracy to 72.13%. By employing Gradient-weighted Class Activation Mapping (Grad-CAM), we empirically reveal that this method generates artifacts that mislead the spatial attention of CNN away from true leukocyte structures. Conversely, adopting Monge-Kantorovich linear color mapping on ResNet-50 architecture effectively preserves semantic structures without introducing false artifacts, which facilitates more localized model attention. This method yields the highest performance, achieving an accuracy of 94.46%, specificity of 98.57%, sensitivity of 94.40% and F1-score of 94.40%, a 3.57% relative improvement over the baseline model. The results underscore the importance of selecting a robust color transfer technique for building reliable automated white blood cell classification systems that remain accurate despite variations in staining protocol.
