Research on melanoma depth of invasion prediction method

Melanoma, a highly malignant skin tumor, relies on its Depth of Invasion (DoI) as a critical metric for assessing tumor malignancy, predicting patient prognosis, and guiding treatment strategies. Traditional DoI measurement methods are manual, time-consuming, and prone to errors due to complex tissue morphologies and the need for fine annotations. This study introduces a novel Convolutional Neural Network (CNN)-based framework that integrates image patch classification with morphological processing to achieve high-precision DoI prediction under coarse annotations. The approach comprises four modules: pathology tissue differentiation using Otsu thresholding and morphological operations, lesion and epidermal region identification via EfficientNetB0 classification, and DoI measurement through least-squares boundary fitting. Experimental results on a melanoma dataset demonstrate a Mean Absolute Error (MAE) of 0.503 mm and a Root Mean Square Error (RMSE) of 0.169 mm, significantly outperforming traditional segmentation networks such as UNet and Attention-UNet. This method provides a robust and efficient solution for automated melanoma diagnosis, with substantial potential for clinical translation.