Adaptive Representation Learning for Prediction and Semantic Segmentation

Doktorsavhandling, 2026

Deep learning models for prediction and segmentation are often designed under idealised assumptions: that inputs are complete, relevant measurements are available, and visual data can be processed at a fixed spatial resolution. Clinical and biomedical applications often violate these assumptions. Observations may be incomplete, heterogeneous, or costly to acquire, and the information needed for prediction or segmentation may be concentrated in a small part of the input.

This thesis develops deep learning methods that use information and computation selectively. For prediction under observation constraints, GenoARM (Paper A) formulates antimicrobial resistance prediction as a joint problem of gene-test selection and resistance classification, using reinforcement learning to construct compact genomic measurement panels. A multi-stream LVEF prediction method (Paper E) estimates cardiac function from echocardiographic examinations where views may be missing, duplicated, or misclassified, combining image sequences and optical flow to handle variable input configurations.

For segmentation, ARTA (Paper B) introduces adaptive mixed-resolution token allocation, concentrating high-resolution tokens in semantically complex regions while keeping simpler regions coarse. BATS (Paper C) extends this principle to volumetric medical segmentation using dense multi-scale allocation and Parent Attention for cross-scale context. SwInception (Paper D) complements these adaptive methods by strengthening local vision Transformers with a multi-scale convolutional structure. Together, the papers show how adaptive representation learning can improve robustness and efficiency under incomplete observations, costly measurements, and spatially uneven visual information.