DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES
M.Tech Research Thesis Defense
Speaker : Mr. Akhil
S.R. Number : 06-18-01-10-22-23-1-23206
Title : “An Adaptive Learning Framework for Brain Tissue Extraction from MRI with Explainability”
Thesis examiner : Prof. Soma Biswas, Department of Electrical Engineering, IISc.
Research Supervisor : Prof. Debnath Pal
Date & Time : June 12, 2026 (Friday) at 11:00 AM
Venue : # 102 CDS Seminar Hall
==========================================================================
ABSTRACT
Brain extraction, or skull-stripping, is a critical first step in nearly every neuroimaging pipeline. When this step fails, the errors propagate downward, degrading the accuracy of volumetric and surface analyses. While neural networks have pushed performance benchmarks higher, they remain computationally expensive, opaque, and often struggle when applied to data from different scanners (domain shift). This thesis investigates whether a carefully designed, fully interpretable classical learning framework can match the accuracy of deep learning while offering better explainability and lower deployment costs. We present a registration-guided, slice-aware framework for extracting the brain from multi-contrast Magnetic resonance imaging (MRI) (T1/T2/FLAIR). The pipeline operates in three stages: (i) it uses registration to characterize patient brain texture and generate reliable training examples; (ii) it employs a gradient boosted classifier to identify the high-confidence core brain tissue; and (iii) it models boundary uncertainty using a gradient-boosted regressor that outputs continuous inclusion scores. The complete feature set, including intensity, rotationally invariant Local Binary Patterns (LBP), Gray-Level Co-occurrence Matrix (GLCM) Haralick features, Gabor responses and wavelets. Empirical evaluation was performed on the Neurofeedback Skull-stripped (NFBS) and Calgary-Campinas-359 (CC-359) datasets, as well as an institutional AIIMS clinical cohort. On NFBS, the proposed method achieved a mean Dice of 0.9521 ± 0.0069 and a mean HD95 of 5.1871 ± 1.5544 mm. On CC359, it attained a mean Dice of 0.9555 ± 0.0073 and a mean HD95 of 4.5633 ± 0.8910 mm. On the AIIMS clinical dataset, it reached a mean Dice of 0.9434 ± 0.0107 and a mean HD95 of 6.4078 ± 0.8750 mm. All values are reported as mean ± per-subject standard deviation. These results demonstrate that this feature-driven pipeline achieves overlap and boundary fidelity comparable to contemporary Convolutional Neural Network (CNN) baselines. Crucially, it does so without requiring large, annotated training corpora and provides fully inspectable decision rules. Qualitative analysis suggests fewer extreme failure cases in challenging slices and more anatomically plausible behavior when the model is uncertain. Current limitations include conservative under-segmentation in regions of complex cortical folding and reduced accuracy in cases of severe anatomical deformation, such as large lesions. Future directions aimed at addressing these limitations include proposals for multi-centre validation and pathology-targeted extensions.
ALL ARE WELCOME
