Integrating Domain Priors in Rule-Based and Deep Learning Algorithms for Biomedical Image Analysis

M.Sc. Trina De

AutorIn

M.Sc. Trina De

Titel

Integrating Domain Priors in Rule-Based and Deep Learning Algorithms for Biomedical Image Analysis

Zitierfähige Url:

https://nbn-resolving.org/urn:nbn:de:bsz:14-qucosa2-1037547

Erstveröffentlichung

2026

Datum der Einreichung

24.10.2025

Datum der Verteidigung

16.03.2026

Abstract (EN)

This work is grounded in the principle that domain-specific priors should be formally encoded as constraints or inductive biases within both the algorithmic design space and the performance evaluation framework. In disciplines such as physics and astrophysics, domain knowledge is routinely formalised via governing equations such as conservation laws or symmetries and integrated directly into analytical or simulation-based models. Similarly, in chemistry, quantum mechanical operators, reaction rate equations, and molecular geometries serve as structured priors that inform model parameterisation and ensure physical consistency. These are often expressed through Hamiltonians, potential functions, or topological invariants embedded in the learning framework. In contrast, biomedical image analysis has traditionally relied on generic computer vision architectures that lack encoding of biologically meaningful constraints, resulting in models that are often agnostic to structural and diagnostic plausibility. This thesis advances a counterpoint by yielding models that are not only computationally effective but also aligned with the underlying biomedical semantics. Contemporary biomedical research and diagnostics are increasingly dependent on microscopy as a primary data acquisition modality. However, the mathematical analysis of such image data presents significant challenges due to the geometric complexity, irregularity of biological structures and modality-specific variability of biomedical imaging. These challenges are compounded by the demand for high-throughput quantitative analysis and the prevalence of spatially nested entities, where structures are hierarchically embedded within one another. This thesis addresses these issues through the design of a biologically constrained model whose architecture, algorithm, objective functions and evaluation criteria are informed by spatial and domain-specific inclusion relations, the development of an algorithmic tool for automated segmentation and quantification, an architecture that leverages local patch representations and additional normalisation to overcome multi-scale detection failures, construction of publicly available, structured and complete microscopy datasets, and an objective function derived from statistical hypothesis testing that can be used in an existing polynomial-based blackbox solver to derive the optimal parameters of a biological process simulator. At first we address the lack of lack encoding of biologically meaningful constraints. For this we develop HydraStarDist with Within Boundary Regularisation Penalty (HSD-WBR), a single-shot, star-convex polygon-based segmentation architecture for biologically nested objects. Here we 1) model the task as a mapping between the image space and segments where each segment is parameterised by a centre point and radial distances describing a star-convex boundary, 2) To enforce biologically realistic containment relationships—such as nuclei residing within cytoplasmic boundaries—we formulate a Within-Boundary Regularisation (WBR) penalty, with can be defined as a inclusion/exclusion/spatial overlap constraint. We also prove here a case of failure for the existing evaluation criteria Average Precision (AP) and subsequently 3) propose a class of new evaluation criteria called Joint True Positive Rate (JTPR) with class relationship specific variants. We show the superiority of our model in the task-relevant evaluation criteria. Next we simultaneously address the lack of of publicly available, structured, complete microscopy datasets and find a way to evaluate the model above in real-world settings. We construct two openly accessible microscopy datasets, namely, the HeLaCytoNuc dataset and the VACVPlaque dataset, capturing spatially nested biological structures— specifically, cytoplasm and nucleus regions within immortalised HeLa (ATCCCCL- 2) cell line and capturing viral plaque phenotypes of Vaccinia virus Western Reserve strain in standard well-plate assays respectively. These datasets serve as testbeds for mathematically grounded segmentation models that bridge discrete object representations with continuous spatial priors while also being publicly available to aid in future task formulation and learning. Further on, we address the problem of an open-source, modular, computational framework for automated segmentation and quantification. Thus, we present PyPlaque, for the automated quantification of viral plaques in fluorescence and digital microscopy images, particularly designed for well-plate assay readouts. Accepting an image as a field of raw intensity values over a spatial domain containing multiple infection foci, the output can be thought of as a structured tuple containing the binary well mask, a collection of disjoint regions of plaques and a feature vector of encoded phenotypic properties such as area, eccentricity, centroid of spread etc. This formalism enables PyPlaque to be modular, interpretable, and extensible, supporting integration with machine learning pipelines and custom assay configurations. As a penultimate step, to circumvent modality-specific and morphometric variability we propose a Patch U-Net architecture for pixel-wise segmentation of cellular content in micrographs imaged using brightfield microscopy. The architecture leverages local patch representations with sparsity-balancing augmentations, content-relevant normalisations and loss functions composed of binary cross-entropy (BCE) and a generalised soft Dice coefficient over the image domain. This model provides a probabilistic inference map, allowing for binary segmentation without reliance on staining or preprocessing, directly from raw clinical images. To note, this also emulates resource-constrained diagnostic setting priors. And in the final chapter, we move away from neural networks as function approximators and a learning paradigm and show a possible methodology of inferring optimal parameters in viral spread such as infection rate, cellular response rates, cellular spread etc. which govern the dynamics of the spread. To achieve this, we define a statistically grounded objective function based on the parametric and non-parametric hypothesis testing frameworks, measuring the dissimilarity between observed and predicted data distributions. Due to the complexity and computational expense of the viral spread simulator, we approximate it with a surrogate model from an existing polynomial-based blackbox solver with coefficients such that it minimise the statistical loss derived above. This surrogate enables tractable optimisation which yields biologically plausible parameters best aligning simulation to empirical measurements. The framework is bidirectionally applicable, since inference on experimental laboratory data yields insight into real infection kinetics and parameter inversion on simulated data demonstrates the flexibility and internal consistency of the simulator by recovering its generative parameters from its own outputs under noise. This unifies simulator validation and real-data interpretation under a common inference framework.

Verweis

Link: http://arxiv.org/abs/2504.12078
Single-shot Star-convex Polygon-based Instance Segmentation for Spatially-correlated Biomedical Objects
DOI: 10.48550/arXiv.2504.12078

A digital photography dataset for Vaccinia Virus plaque quantification using Deep Learning
DOI: 10.1038/s41597-025-05030-8

PyPlaque is an open-source python package for phenotypic analysis of virus plaque assays
DOI: 10.1038/s41598-025-20075-w

A clinical microscopy dataset to develop a deep learning diagnostic test for urinary tract infection
DOI: 10.1038/s41597-024-02975-0

Forschungsdatenverweis

HeLaCytoNuc: fluorescence microscopy dataset with segmentation masks for cell nuclei and cytoplasm
DOI: 10.14278/rodare.3001
Link: https://doi.org/10.14278/rodare.3001

VACVPlaque: mobile photography of Vaccinia virus plaque assay with segmentation masks
DOI: 10.14278/rodare.3003
Link: https://doi.org/10.14278/rodare.3003

A Dataset for Virus Infection Reporter Virtual Staining in Fluorescence and Brightfield Microscopy
DOI: 10.14278/rodare.3742
Link: https://doi.org/10.14278/rodare.3742

Clinical urine microscopy for urinary tract infections
DOI: 10.14278/rodare.2473
Link: https://doi.org/10.14278/rodare.2473

Freie Schlagwörter (DE)

Domänenprioren, Bildanalysealgorithmen, Deep Learning, biomedizinische Bildgebung

Freie Schlagwörter (EN)

Domain Priors, Image Analysis Algorithms, Deep Learning, Biomedical Imaging

Klassifikation (DDC)

006

Klassifikation (RVK)

ST 330

ST 301

WC 7722

GutachterIn

Prof. Dr. Ivo F. Sbalzarini
Prof. Dr. Michael Hecht

Den akademischen Grad verleihende / prüfende Institution

Technische Universität Dresden, Dresden

Förder- / Projektangaben

Center for Advanced Systems Understanding

Bundesministerium für Forschung, Technologie und Raumfahrt and Sächsische Staatsministerium für Wissenschaft, Kultur und Tourismus Center of Excellence for AI-research “Center for Scalable Data Analytics and Artificial Intelligence Dresden/Leipzig”
ID: ScaDS.AI, DEAL

HelmholtzAI ID: tomoCAT

Sonstige beteiligte Institution