Spatial Features-Oriented Semantic Segmentation of Tumour Regions in Breast Histopathology Images

Yew, Jee Shem (2025) Spatial Features-Oriented Semantic Segmentation of Tumour Regions in Breast Histopathology Images. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text Yew Jee Shem - FULL TEXT.pdf Restricted to Registered users only Download (12MB)

Abstract

Accurate delineation of tumour regions in histopathology images of breast tissue is crucial for improving diagnostic precision and guiding treatment decisions. This study presents a technical framework that integrates advanced pre-processing, image enhancement, clustering and spatial refinement into a unified semantic segmentation pipeline. Initially, twenty-five regions of interest were extracted from three anonymised breast histopathology slides scanned with an Aperio CS2 whole-slide scanner at 10x magnification, resulting in 24-bit RGB images of 583 x 1276 pixels at 0.2521 μm/pixel. A two-stage pre-processing workflow first identifies red artefacts arising from residual blood and converts affected pixels to representative cytoplasm colours using threshold analysis across red, green and blue channels. Global colour profiles are then standardised via Macenko stain normalisation to mitigate slide-to-slide staining variability. Local contrast and tissue detail are further enhanced by applying contrast limited adaptive histogram equalisation to the luminance channel without introducing spurious boundaries. Enhanced images are converted to the Lab colour model and each pixel is represented by a four-dimensional feature vector comprising the a* and b* chromatic components coupled with scaled x and y coordinates. This joint encoding of chromatic and spatial information empowers the segmentation algorithm to resolve spatially ambiguous tumour boundaries and subtle texture variations. The optimal number of segmentation classes is determined by locating the elbow point in the plot of within-cluster variance as cluster count increases. A K-Means algorithm partitions pixels into background, non-tumour, and tumour clusters according to their feature vectors and mean grayscale intensities. To improve anatomical coherence and remove isolated misclassifications, a proposed Spatial Neighborhood Superpixel Refinement (SNSR) module is applied. The proposed SNSR generates superpixels based on local feature contrast and uses a soft voting rule to relabel entire regions when a specified fraction of pixels is initially classified as a tumour. Constrained hole filling then merges small gaps within tumour areas to produce contiguous masks aligned with natural tissue structures. The final binary tumour masks are compared to manually annotated ground truth for twenty-five images drawn from slides labelled B17, B2, and B36. Quantitative evaluation metrics include overall Accuracy of 93.39%, Sensitivity of 90.17%, Specificity of 94.44%, Precision of 90.12%, and F1 Score of 90.11%, demonstrating a balanced high performance. These results confirm the method reliability of the method in distinguishing tumours from non-tumour regions. This framework establishes a solid technical basis for future automated breast cancer histopathology analysis.