DISSERTATION DEFENSE

Author :  Hongpeng Yang
Advisors: Dr. Yan Tong
Date: Aug 11, 2026
Time: 02:00 pm
Location: Meeting room 2267
Link:  https://teams.microsoft.com/meet/216585525604914?p=nTgWeDedV9AAIGAdFn

Abstract

Deep learning models for medical image segmentation are typically trained under assumptions that clinical data do not always satisfy: distinguishable anatomical boundaries, complete volumetric observations, and access to every required imaging modality. In practice, lesion contrast may be weak, slices may be missing along the depth axis, or MRI sequences may be unavailable. Each condition removes a different form of evidence and calls for tailored compensation. This dissertation develops three task-oriented frameworks for such structured input imperfections, that is, imperfections whose location or identity is known before the prediction is made. Although addressing distinct problems, all three adapt internal representations to the affected information and keep compensation tied to segmentation.

The first framework, Dark Vision Network (DVNet), addresses weak boundary and texture cues in U-shaped networks, by analogy with perception under low illumination. It decomposes skip-connection features into frequency subbands and uses Mamba-based, contrast-oriented fusion to enhance them before decoding. Its plug-and-play design leaves the backbone unchanged and applies to CNN-, Transformer-, and Mamba-based architectures. Across these backbone families, the module lowers boundary distance error in every reported comparison and improves region overlap in most of them, and the same design transfers from volumetric brain MRI to abdominal MRI and two-dimensional microscopy.

The second framework, InterFrameNet, addresses spatial incompleteness when only the endpoint slices of a local window are observed. Rather than reconstructing missing images, it predicts intermediate segmentation features from endpoint features and relative slice positions. An auxiliary delta-aware objective regularises structural changes between adjacent predictions. Experiments under increasing slice sparsity show that segmentation-directed cross-frame prediction outperforms copy- and mean-based filling, and that its advantage widens as the gap between observed slices grows.

The third framework FAR-Seg, addresses missing modalities in brain tumour segmentation. It estimates a target-specific feature for each absent MRI sequence while preserving the acquired-modality features. A predicted class-wise failure map then guides residual correction of the initial segmentation. Evaluation over every non-empty modality combination shows that separating feature completion from output correction improves segmentation accuracy, with the clearest gains where the absent sequence carries evidence specific to one tumour region, such as enhancing tumour when contrast-enhanced T1 is unavailable.

Together, these frameworks support a common design principle for medical image segmentation under imperfect inputs: compensation should reflect the degraded or missing evidence and operate at the stage most directly connected to segmentation. This principle links feature enhancement, cross-frame prediction, and feature completion with output correction within a single task-oriented perspective.