Seyed Hamid Khatami, Nasser Mehrshad, Seyed Hamid Zahiri
This study demonstrates that a simple, perception-inspired algorithm can effectively address the challenges of breast tumor segmentation, achieving accuracy and efficiency comparable to those of complex deep neural networks, while offering decisive advantages in interpretability, speed, and independence from annotated training data. Our findings support the design of training-free diagnostic aids that are particularly well-suited for resource-constrained clinical environments, where large datasets and advanced GPUs are not readily available. This research emphasizes the importance of designing algorithms that replicate the systematic diagnostic processes used by human experts.
BACKGROUND: The development of deep learning techniques has greatly improved tumor detection and analysis in breast ultrasound images. Despite their impressive performance, deep networks often face several challenges, including high computational complexity, large model sizes, limited generalization, and reliance on large labeled datasets. This research presents a simple, training‑free method that achieves highly competitive accuracy compared to state‑of‑the‑art deep learning models, while eliminating the need for large annotated datasets, high‑performance GPUs, and lengthy training procedures.
METHODS: We present a perception-inspired hierarchical pruning method that aims to accurately detect tumor regions in medical images by removing non-tumor regions based on the statistical distribution of intensities. The automatic pruning stops when the mean and median intensities of the remaining pixels converge, and a candidate tumor region remains.
RESULTS: The proposed algorithm demonstrated excellent performance on both the BUSI and UDAIT datasets. On the BUSI dataset, it achieved an accuracy of 99%, a sensitivity of 93%, a specificity of 99%, and an F1-score of 94%. Similarly, on the UDAIT dataset, it reached an accuracy of 99%, a sensitivity of 91%, a specificity of 97%, and an F1-score of 92%. These results highlight the algorithm's strong ability to minimize both false positive and false negative rates, offering performance comparable to that of expert-level analysis.
CONCLUSION: This study demonstrates that a simple, perception-inspired algorithm can effectively address the challenges of breast tumor segmentation, achieving accuracy and efficiency comparable to those of complex deep neural networks, while offering decisive advantages in interpretability, speed, and independence from annotated training data. Our findings support the design of training-free diagnostic aids that are particularly well-suited for resource-constrained clinical environments, where large datasets and advanced GPUs are not readily available. This research emphasizes the importance of designing algorithms that replicate the systematic diagnostic processes used by human experts.