CBi-BERT: Efficient Skin Disease Image Segmentation Using Patch-Based Deep Feature Mapping and BERT-Based Attention Mechanism

 

Summi Goindi1,*, Khushal Thakur2, Divneet Singh Kapoor2

1Research Scholar, Chandigarh University, Mohali, Punjab

2Associate Professor, Chandigarh University, Mohali, Punjab

Text Box: Abstract Skin image segmentation serves as a vital undertaking in medical image analysis, specifically in dermatology, since it enables the detection of skin diseases and the assessment of effectiveness of treatments. Segmenting skin lesions from photographs is a crucial step in working towards this patchive. Nevertheless, the work of segmenting skin lesions is difficult due to the existence of both artificial and natural deviations, inherent characteristics like the shape of the lesion), and deviations in the circumstances during which the images are obtained. In recent years, researchers have been investigating the feasibility of utilizing deep-learning models for skin lesion segmentation. Deep learning methodologies have exhibited encouraging outcomes in various image segmentation initiatives, thereby preventing the possibility of automating and enhancing the precision of skin segmentation. This paper introduces a new hybrid method, named the CBi-BERT framework, aimed to improve the results and architectures of medical image segmentation or patch detection tasks. This architecture employs Convolutional Neural Networks (CNNs) for feature extraction as well Bidirectional LSTM-based encoders to process sequence information and BERT based attention collection across the strongest features. Image normalization, resizing and data augmentation techniques are applied in the proposed method to deal with major challenges faced during medical imaging such as rate of image quality differentiation from noise or bias between benign vs. malign features. We evaluate the performance of CBi-BERT to those benchmark datasets and state-of-the-art models, showing that CBi-BERT outperforms them in all relevant metrics such as Intersection over Union (IoU), recall, mean average precision (bin-MAP) DICE coefficient. Specifically, for images sized 256x256 the model achieved IoU =0.9, recall=0.92, mAP=0.89 and Dice coefficient: =0.91 thereby outperforming some advanced state-of-the-art models as ResNet50, VGG16, UNet, EfficientNet-B-01 Our results show that the framework is able to detect and segment important structures in medical images with high precision which makes it a compelling tool for AI based Healthcare solutions. Emails: summikhurana7@gmail.com; khushal.thakur@cumail.in; divneet.ece@cumail.in

Received: January 28, 2025 Revised: March 03, 2025 Accepted: April 22, 2025

 Keywords: Skin; Image segmentation; Deep Learning; IOU; Segmentation