Bilkent University
Department of Computer Engineering
M.S.THESIS PRESENTATION
TOPOLOGY-PRESERVING MEDICAL IMAGE SEGMENTATION VIA PATH-COST SUPERVISION
Elyar Esmaeilzadeh
Master Student
(Supervisor: Asst.Prof.Doruk Öner)
Computer Engineering Department
Bilkent University
Abstract: Curvilinear and surface-like anatomical structures must be segmented accurately for quantitative work in connectomics, vascular imaging, and cell biology. Networks trained with pixel-wise losses often reach high overlap scores yet leave small topological errors, such as broken branches, spurious connections, or holes, that harm downstream measurements. This thesis proposes two loss functions that supervise connectivity by penalizing shortest-path traversal costs on predicted distance maps. CAPE (Connectivity-Aware Path Enforcement) builds a ground-truth graph from centerline annotations and compares shortest paths between connected vertex pairs with paths on the network prediction. Squared distance values along predicted paths and a masking strategy that blocks alternate loop routes yield dense gradients along trajectories and direct connectivity optimization. On 2D neuron and retinal vessel datasets and a 3D mouse brain volume, CAPE achieves the best reported APLS and TLTS among the compared methods while keeping competitive pixel-wise performance. GEPA (Generalized Path Cost Loss) extends the path-cost idea to dense supervision for curvilinear and surface structures. Where prediction and ground truth disagree, it finds connected components, selects a consensus source per component, and compares single-source shortest-path cost maps from Dijkstra’s algorithm. Pathfinding cost drops from O(K·N logN) to O(N logN) per component; graph extraction is not required. GEPA penalizes false disconnections and false connections across whole regions. On seven benchmarks, including neurons, retinal vessels, the Circle of Willis, cellular surfaces, and organelle segmentations, it reduces Betti matching error and improves connectivity scores with modest training overhead.
DATE: June 22, Monday @ 13:00
Place: Zoom
Zoom link: https://zoom.us/j/96023681106?pwd=D3rCF7CYyo9tz41HwcYHO2e2dTwZgc.1
Meeting ID: 960 2368 1106
Passcode: 896289