Automatic Estimation of Region of Interest Area in Dermatological Images Using Deep Learning and Pixel-Based Methods: A Case Study on Wound Area Assessment

Accurate wound area estimation is essential for effective dermatological assessment and treatment monitoring. However, manual measurement is time-consuming and error-prone, highlighting the need for automated, reliable methods. This paper aims to develop and evaluate two complementary techniques for estimating the Region of Interest (ROI) in dermatological images: a novel deep learning approach using the Segment Anything Model (SAM) and a simple pixel-based thresholding method. SAM segments both the wound and a reference object automatically or through prompt-based queries, without requiring additional supervised classification. The pixel-based method offers a lightweight alternative for resource-limited settings. Both techniques generate binary masks and calculate real-world areas using a pixel-to-centimeter scale. Evaluation on 40 images shows that SAM outperforms the pixel-based method, achieving an average relative error of 4.63% versus 9.5% and ≤5% error in 62.5% of cases compared to 27.5%. The proposed methods are not limited to wound area estimation but can be extended to inflammation area detection in rheumatoid arthritis and ophthalmology, providing a scalable framework for ROI estimation in medical imaging.