Integrating Median Filtering with Deep Pattern Recognition Models for Improved Breast Ultrasound Cancer Detection

Abstract

Even though ultrasound of the breast is an easy to access imaging modality and is radiation free, speckle-like intensity fluctuations, low contrast, variability of acquisition and limited model transparency are the challenges in automated classification of ultrasound lesions. In this paper, we introduce a controlled framework of preprocessing-classification-explainability that compares the aforementioned three image types (raw, 3 × 3 median filtered, median filtered + contrast-limited adaptive histogram equalization (CLAHE)) using the same leakage-safe data split, convolutional neural network (CNN), training budget and evaluation protocol. To minimize the information leakage, the downloaded Kaggle archive of benign and malignant images was pre-augmented with 9016 images; and then divided into 70/15/15 train/validation/test sets based on augmentation-family grouping. The 96.97% accuracy, 98.95% sensitivity, 95.04% specificity, 95.10% precision, 96.99% F1-score, 0.99623 ROC-AUC and 0.99608 PR-AUC of the model that was median-filtered were the highest. Compared to raw standardized baseline, median filtering resulted in an increase in accuracy and sensitivity by 3.55 and 6.15 percentage points, respectively, and in a decrease of false negative cases from 48 to 7. Median restoration also gave the best results in terms of PSNR and SSIM for controlled synthetic-speckle testing, which were 27.9313 dB PSNR and 0.8214 SSIM, respectively. There was a significant reduction in paired errors after raw versus median filtering (adjusted p = 0.000008) by Exact McNemar analysis. Grad-CAM, LIME and SHAP also revealed class-specific areas and failure modes but only showed low spatial overlap, which indicated that these are complementary rather than interchangeable. The results indicate that a very basic median filter which preserves the edges can enhance the predictive reliability and model auditing when the effects of the pre-processing are separated in an experimental setup.