Filters and Learning-based Methods for Tumor Detection from Ultrasonic Images

Abstract

None

Tumor segmentation in medical imaging plays a critical role in the accurate diagnosis and treatment planning of cancer. This study proposes a hybrid framework that combines complementary convolutional neural network (CNN) models and advanced post-processing techniques to achieve robust and accurate tumor segmentation. The initial model (Model 1) employs CLAHE preprocessing, CNN-based predictions, and active contour refinement to provide a baseline segmentation. However, its performance is limited by difficulties in capturing complex tumor boundaries. To address these challenges, a second model (Model 2) incorporates noise-augmented preprocessing and iterative detection, enhancing the segmentation of subtle and irregular tumor regions. The outputs of both models are merged using logical operations and refined further with edge correction and size filtering. Additionally, an enhanced merging model integrates a Spatial Intensity Metric (SIM) expansion, which leverages spatial and intensity relationships to refine and expand tumor regions, particularly addressing under-segmented areas. This enhancement results in significant improvements, as demonstrated by higher F1 and IoU scores compared to earlier models. The study also highlights the limitations of the grid-based 16×16 classification approach, especially for large tumors, and suggests future directions such as adaptive grid sizes, more detailed labeling schemes, and the incorporation of local texture analysis for malignancy assessment. The proposed framework demonstrates the potential of integrating machine learning and traditional image processing techniques for accurate tumor segmentation, paving the way for more reliable and clinically valuable diagnostic tools.