Automatic Acetowhite Lesion Segmentation via Specular Reflection Removal and Deep Attention Network

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Pathology

Background:

• Accurate segmentation of acetowhite lesions in colposcopy images (cervigrams) is crucial for diagnosing cervical intraepithelial neoplasia and cervical cancer.
• Existing computer-aided diagnosis algorithms struggle with segmentation due to specular reflections, limited data, and difficulty focusing on relevant lesion areas.

Purpose of the Study:

• To develop a novel computer-aided diagnosis algorithm for automatic and accurate segmentation of acetowhite lesions in cervigrams.
• To enhance diagnostic accuracy and pathological examination guidance for gynecologists.

Main Methods:

• A specular reflection removal mechanism was developed to precisely detect and inpaint reflective areas in cervigrams.
• A cervigram image classification network was designed to predict pathology results and generate lesion attention maps.
• A lesion-aware convolutional neural network utilized attention maps to guide precise acetowhite lesion segmentation.

Main Results:

• The proposed method demonstrated superior performance in segmenting acetowhite lesions compared to state-of-the-art approaches.
• The algorithm achieved improved Dice similarity coefficient and Hausdorff Distance values on a dataset of 3045 clinical cervigrams.
• The specular reflection removal and attention-guided segmentation significantly enhanced segmentation accuracy.

Conclusions:

• The novel algorithm effectively addresses challenges in cervigram segmentation, including specular reflections and focus on lesion areas.
• The developed method offers a significant advancement in computer-aided diagnosis for cervical cancer screening.
• This approach has the potential to improve the accuracy and efficiency of pathological examination guidance in colposcopy.