Efficient fine-tuning of deep learning CNN models with the level set method for breast cancer ultrasound image segmentation

Breast cancer is the most frequently diagnosed malignancy among women and remains a major cause of cancer-related mortality worldwide. Ultrasound is widely used as an adjunct to mammography, particularly in women with dense breast tissue, but speckle noise, acoustic artefacts, and poorly defined lesion margins make automated segmentation extremely challenging. Reliable delineation of breast lesions in ultrasound images is nonetheless essential for quantitative assessment, radiomics analysis, and computer-aided diagnosis.This dissertation develops and evaluates a hybrid breast ultrasound segmentation framework that couples convolutional neural networks (CNNs) with Distance Regularized Level Set Evolution (DRLSE). Two public datasets, BUSI and BUS-UCLM, are first preprocessed by extracting lesion-centred regions of interest (ROIs) at three spatial scales (large, medium, and small) and resampling them to a standard $128\times 128$ format with intensity normalisation. Lesion shape is quantified using a circularity index computed from the ground-truth masks, and each case is assigned to one of three difficulty levels (easy, medium, or hard). This circularity-based stratification is used both to analyse segmentation performance as a function of morphological complexity and to highlight cases that are intrinsically difficult for any algorithm.In the proposed pipeline, previously fine-tuned CNN segmentation models (AlexNet, U-Net, VGG19, ResNet18, ResNet50, MobileNet, and Xception) are applied to the ROIs to generate coarse lesion probability maps and corresponding binary masks. These CNN masks serve as signed-distance initialisations for an edge-based DRLSE model, which enforces geometric regularity and drives the evolving contour toward image-derived edge features. A systematic comparison is conducted between (i) the purely CNN-based segmentations, (ii) a classical DRLSE method initialised without deep learning, and (iii) the CNN–DRLSE hybrid configuration. Performance is assessed using standard overlap and boundary metrics, including Dice similarity coefficient, intersection over union, precision, and recall.Experiments on BUSI and BUS-UCLM show that all CNN baselines substantially outperform standalone DRLSE in terms of Dice and IoU across ROI sizes and difficulty levels, whereas DRLSE alone tends to under-segment lesions but yields smooth contours. Reducing the ROI around the lesion consistently improves CNN performance, demonstrating the benefit of focused, standardised crops. Segmentation accuracy decreases monotonically from easy to hard circularity classes for all methods, confirming that irregular, low-circularity lesions are more challenging. The hybrid CNN–DRLSE approach preserves the strong overlap scores of the CNNs while removing small spurious regions, filling internal holes, and producing more anatomically plausible boundaries, with modest but meaningful improvements in difficult BUS-UCLM cases. Each hybrid inference requires less than one second on commodity hardware, indicating that the method is computationally efficient and compatible with near real-time clinical workflows. Overall, the results demonstrate that integrating pretrained CNNs with DRLSE and circularity-based lesion stratification yields a robust, interpretable, and practically deployable framework for breast ultrasound segmentation.