arXiv subjects

Karl Thomas Hjelmervik

Publications and source records attributed to Karl Thomas Hjelmervik.

3 recordsLinked to original sources

arXiv2026

Decomposing Whole Slide Image Report Generation with Graph-Constrained Multiple Instance Learning Workflows

Whole-slide image (WSI) report generation requires recognizing spatially distributed pathological features and organizing them into a coherent diagnostic narrative. Although direct vision-to-text models can yield fluent reports, they obscure the contributions and failure modes of visual recognition, structured reasoning, and language generation. We propose a decomposed framework in which frozen Virchow2 tile embeddings are aggregated by multiple-instance learning (MIL) classification heads that answer organ-specific diagnostic questions. An organ-conditioned graph constrains the assembly of these answers into a structured reasoning chain, which a language model realizes as a pathology report. On the REG2026 held-out set of 2,028 slides, the proposed workflow achieved a chain-Jaccard score of 0.702. Performance fell to 0.420 without graph-based chain construction, 0.398 when the organ-specific graphs were replaced by a single organ-agnostic graph, and 0.371 when the language model constructed the chain freely from MIL predictions. Using the same report generator, graph-structured chains improved the report score from 0.330 to 0.495. On 350 external TCGA WSIs spanning the seven REG organs without fine-tuning, the expected organ graph was selected in 64.0% of cases and ranked among the top three in 86.6%. Providing the correct organ graph increased agreement with coarse TCGA primary-diagnosis labels from 61.8% to 92.6%, identifying organ routing as a main bottleneck under domain shift. Overall, organ-conditioned, graph-constrained chain assembly improves structured reasoning and report generation while enabling stage-specific error localization.

cs.CV↗

arXiv2026

Multi-Stage Fine-Tuning of Pathology Foundation Models with Head-Diverse Ensembling for White Blood Cell Classification

The classification of white blood cells (WBCs) from peripheral blood smears is critical for the diagnosis of leukemia. However, automated approaches still struggle due to challenges including class imbalance, domain shift, and morphological continuum confusion, where adjacent maturation stages exhibit subtle, overlapping features. We present a multi-stage fine-tuning methodology for 13-class WBC classification in the WBCBench 2026 Challenge (ISBI 2026). Our best-performing model is a fine-tuned DINOBloom-base, on which we train multiple classifier head families (linear, cosine, and multilayer perceptron (MLP)). The cosine head performed best on the mature granulocyte boundary (Band neutrophil (BNE) F1 = 0.470), the linear head on more immature granulocyte classes (Metamyelocyte (MMY) F1 = 0.585), and the MLP head on the most immature granulocyte (Promyelocyte (PMY) F1 = 0.733), revealing class-specific specialization. Based on this specialization, we construct a head-diverse ensemble, where the MLP head acts as the primary predictor, and its predictions within the four predefined confusion pairs are replaced only when two other head families agree. We further show that cases consistently misclassified by all models are substantially enriched for probable labeling errors or inherent morphological ambiguity.

cs.CV↗

arXiv2024

Inverting the sound speed profile from multi-beam echo sounder data and historical measurements -- a simulation study

The ocean's opacity poses challenges for security, as new technology, e.g. underwater drones, offers new opportunities for illegal activities, such as smuggling and terrorism. A network of unmanned surface vehicles (USV) and autonomous underwater vehicles (AUV) offers a potential underwater surveillance solution, but demands high autonomy and compact hardware. For improved situational awareness and efficient operation, sonar performance models may provide the network with sensor coverage maps, but this requires constantly updated environmental information, in particular the present sound speed profile (SSP). We propose the inversion of SSPs from multibeam echo sounder (MBES) data in an environment with known topography. The method exploits the two-way travel time from the MBES to the bottom, comparing the measurements to modelled travel time for a proposed SSP model. An acoustic raytracer models the travel time for the SSP model. The inversion problem is shown to be non-unique when basing the cost function on the two-way travel time alone. This is resolved by incorporating a Tikhonov-type regularization term for inclusion of a priori knowledge on the SSPs in addition to the travel time in the final cost function. Empirical orthogonal functions (EOFs) are derived from a historical SSP data set, and variance for the EOF coefficients are determined from the same data set. The EOF coefficient distributions are assumed Gaussian and used in the regularization term to limit the search space of the inversion algorithm to physically feasible SSPs. A neural network determines the regularization parameters. The method's validity and sensitivity to errors is assessed using synthetic sonar data for the Norwegian Trench. The method accurately recovers SSPs with average root mean square errors of 0.83 m/s. For comparison, the error obtained using state-of-the-art climatology (WOA) is 2.6 m/s.

eess.SP↗