Enhancing HIFU Lesion Area Detection through Supervised and Contrastive Self-Supervised Learning with Wavelet-Based Feature Extraction and Hard Negatives (HWCSSL)

Introduction: Artificial intelligence (AI) has brought about transformative changes across multiple domains including classification, detection, and prediction through the implementation of sophisticated machine learning algorithms, with the ultimate objective of enhancing quality of life and service delivery. Conventional approaches that rely on manually engineered features encounter significant limitations when applied to complex tasks such as medical diagnosis, particularly in terms of suboptimal feature extraction capabilities and heightened sensitivity to noise interference. The emergence of deep neural networks, with their inherent capacity for automated feature extraction, has revolutionized contemporary data analysis methodologies. The current study leverages this technological advancement for the detection of lesions generated by High-Intensity Focused Ultrasound (HIFU), an innovative therapeutic technique developed for oncological treatment and hemorrhage management. The proposed methodology integrates both supervised and self-supervised learning paradigms to enhance detection accuracy by effectively utilizing both labeled and unlabeled clinical datasets. Method: A key challenge in traditional methods lies in optimal feature extraction and hyperparameter tuning. This research develops an innovative framework integrating supervised and contrastive self-supervised learning. The model processes both RF signals and B-mode images as inputs, simultaneously handling labeled and unlabeled data. Data augmentation techniques, including wavelet transforms and hard negative sampling were employed. Model optimization was achieved through advanced algorithms and fine-tuned hyperparameters to enhance performance in complex scenarios. This approach enabled concurrent analysis of multimodal data, improving diagnostic accuracy. Results: The proposed model demonstrated substantial improvements in HIFU-induced lesion classification performance. Quantitative evaluation metrics, including accuracy, precision, recall, and F1-score, consistently validated the model's efficacy in discriminating between healthy and pathological tissue regions. The multimodal integration of signal processing and image analysis components yielded remarkable enhancements in overall system performance. The incorporation of self-supervised learning algorithms coupled with wavelet transform techniques substantially augmented the model's feature extraction capabilities, resulting in superior diagnostic accuracy compared to conventional approaches. Conclusion: This investigation presents a robust and precise framework for HIFU lesion detection that effectively employs self-supervised learning methodologies. The developed system successfully extracts clinically relevant information from unlabeled datasets while achieving significant improvements in diagnostic reliability. By addressing critical challenges in medical imaging applications, particularly in non-invasive therapeutic interventions, this approach demonstrates considerable potential for enhancing clinical workflows. The proposed methodology establishes a foundational platform for the future development of advanced diagnostic tools in therapeutic ultrasound applications.