Abstract

The serious risks to public health from the rapid spread of misleading information brought on by our increasing reliance on digital healthcare information. Recognizing such misinformation is essential in the medical industry to ensure that accurate and trustworthy information is disseminated. We suggest an ideal deep learning-based model combining AnatoFormer and LREN-MedNet for ROI-Free Breast Cancer Diagnosis using self-supervised learning and anatomical-aware feature extraction. Our suggested AnatoFormer- LREN system uses Masked Autoencoder (MAE) pretraining and Contrastive Learning (MoCo) to enhance feature extraction from unlabeled ultrasonic images. The AnatoFormer design improves diagnostic accuracy and model interpretability by accurately capturing intra-layer and inter-layer spatial interactions in breast ultrasound images. Because the resulting representations are employed in a fully automated pathway, eliminating the need for manual ROI annotation, the model is more practical for clinical applications. Additionally, optimization techniques like self-supervised training and fine-tuning methodologies have been employed to increase the resilience of our model. The models performance was assessed on the BUSI dataset and it performs better in classification than transformer-based and traditional CNNs. The suggested AnatoFormer-LREN structure offers a new understandable and effective alternative to automated breast cancer diagnosis advancing artificial intelligence in clinical decision- making and medical imaging.

Authors

D. Monica Seles, Prahanya Selvakumar, M. Kawena
Mepco Schlenk Engineering College, India

Keywords

Breast Cancer Diagnosis, AnatoFormer, LREN-MedNet, Self- Supervised Learning, Contrastive Learning, Masked Autoencoder (MAE), Ultrasound Imaging, Deep Learning, ROI-Free Diagnosis, Medical Image Analysis, Clinical Decision Support