Comparison of Segmentation Techniques for Enhancing LIME-Based Explainability

Ho, Wen Han (2026) Comparison of Segmentation Techniques for Enhancing LIME-Based Explainability. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text HO WEN HAN_Full Text.pdf Restricted to Registered users only Download (8MB)

Abstract

Convolutional neural networks (CNNs) can classify brain-tumour MRI but are difficult to interpret. LIME offers case-wise explanations that depend on how images are partitioned into superpixels. This thesis investigates that dependency by implementing a VGG16 classifier and systematically comparing three segmentation methods Quick Shift, Felzenszwalb, and SLIC via one-factor-at-a-time parameter sweeps around fixed baselines. A slice-level MRI dataset is processed with consistent preprocessing; classification is reported using accuracy, precision, recall, and F1, while explanation behaviour is profiled by surrogate fidelity to the CNN, superpixel granularity/shape, segmentation and explanation time, qualitative overlays, and robustness to small parameter and seed changes. The trained model attains 95.5% test accuracy with strong macro averages. Comparative results show that the choice of segmenter and its granularity materially shape LIME maps, SLIC yields fast, tidy overlays, Felzenszwalb aligns more closely with anatomical boundaries, Quick Shift captures fine detail but is parameter-sensitive and can emphasise strong non-tumour edges. Across methods, mid-range segment counts balance fragmentation and over-merging, and increasing LIME samples stabilises fidelity with diminishing returns. The study summarises these findings into a concise, practice-oriented guide for selecting segmentation settings in MRI explanation workflows