Manuscript received November 27, 2024; revised December 25, 2024; accepted March 25, 2025; published May 20, 2025.

Abstract—Colposcopy, a medical procedure for diagnosing and treating cervical cancer, generates medical images that often contain Specular Reflections (SRs). These bright areas, caused by moisture and device lighting, can affect further image analysis steps like feature extraction and classification. In this work, a two-stage U-shaped convolutional neural network architecture (UNET) diffusion model is employed for SR detection and inpainting in colposcopy images. The first stage creates SR region masks using local thresholding and top-hat morphological operations on augmented and pre-processed images. The second stage employs a UNET diffusion model trained with Dreambooth, and inpaints Specular Reflection (SR) regions from the processed images. The International Agency for Research on Cancer dataset utilized in the overall work comprises 913 colposcopy images to train and evaluate the developed model. The UNET diffusion model achieved reconstruction with a minimal loss of 0.169 and a stable Structural Similarity Index Measure (SSIM) of 0.85. State-of-the-art methods for specular reflection detection and inpainting in colposcopy images evaluated on limited datasets, often lacking diversity in imaging conditions. In contrast, the presented approach is developed to work on colposcopy images taken under varying conditions, including green filter, iodine staining, and acetic acid application, which is required for accurate diagnosis. This enhances the model’s robustness, enabling it to perform effectively across a diverse range of colposcopy images.

Keywords—colposcopy, specular reflection, Unet diffusion

Cite: Parimala Tamang, Annet Thatal, Mousumi Gupta1, and Snehashish Bhattacharjee, "Development of Preprocessing Stage for Early Cervical Cancer Detection Using UNET Diffusion Model," Journal of Image and Graphics, Vol. 13, No. 3, pp. 245-252, 2025.

Copyright © 2025 by the authors. This is an open access article distributed under the Creative Commons Attribution License (CC-BY-4.0), which permits use, distribution and reproduction in any medium, provided that the article is properly cited, the use is non-commercial and no modifications or adaptations are made.