Breast tissue classification using modified Otsu thresholding and disk structuring elements

Yasiran, Siti Salmah and Abdullah, Abdul Muhaymin and Abd Razak, Muhammad Hasif and Abd Halim, Suhaila and Azman, Nor Farihah (2026) Breast tissue classification using modified Otsu thresholding and disk structuring elements. Mathematics Letters, 5 (2): 3. pp. 17-28. ISSN eISSN: 2948-3735
Abstract

Breast tissue density is an important factor in mammographic interpretation, and a computer-aided diagnosis (CAD) pipeline can support the classification of fatty, glandular and dense tissue. This study compares two segmentation pipelines: hybrid original Otsu thresholding with mathematical morphology and hybrid modified Otsu thresholding with mathematical morphology. Ninety-five mammographic images from the Mammographic Image Analysis Society (MIAS) database were used. The images were preprocessed using median filtering and histogram stretching, segmented using the two Otsu-based methods, and refined by erosion with disk-shaped structuring elements having radii from 20 to 200 pixels. Binary Robust Invariant Scalable Keypoints (BRISK) were used for feature extraction and detection, while Principal Component Analysis (PCA) with 95% explained variance was evaluated for feature reduction. Support Vector Machine (SVM) classifiers with Gaussian, linear, quadratic, and cubic kernels were then compared using a 70% training and 30% testing split. On the testing data, the hybrid original Otsu pipeline at radius 200 achieved 88.9% accuracy without PCA and 87.1% with PCA, whereas the hybrid modified Otsu pipeline at radius 80 achieved 87.1% without PCA and 85.2% with PCA. The original Otsu hybrid therefore produced the highest reported testing accuracy, while the results also demonstrate that the disk radius affects classification performance and should be selected empirically for the image set.

Item Details
Edit Item
Edit Item
Downloads & Files
[thumbnail of 149384.pdf]
Text
149384.pdf
Download (586kB)
Indexing & Metrics