Imaging-based spatial transcriptomics (SRT) is an emerging family of technologies that enable transcript-level molecular profiling at subcellular resolution, unlocking new insights into tissue organization, cell-cell interactions and microenvironmental structure. However, the dominant analytical frameworks in the field often rely on segmentation-based paradigms, often borrowed from established single-cell analysis, which arguably introduces inappropriate assumptions and underutilizes the spatial information present in the data. This thesis contributes to the development of a segmentation-free analysis framework for spatial transcriptomics, proposing three dedicated computational tools that operate directly on transcriptomic signal without the need to aggregate observed signal into cell segments: SSAM-lite addresses the need for accessible, segmentation-free cell typing. Building on the kernel density estimation (KDE) approach of the SSAM framework, SSAM-lite provides a performance-optimized, browser-based implementation that enables users to infer cell types directly from transcript coordinates. The method applies spatial smoothing and marker gene-based cell typing in a user-friendly, interactive interface, facilitating exploratory analysis of imaging-based SRT datasets. Validation on benchmark tissue slices shows that SSAM lite replicates expert-annotated cell type maps with high fidelity while lowering technical barriers to entry. Sonar introduces a lattice-based spatial statistics framework for analyzing co-occurrence relationships between feature categories in pixelized cell type maps. The method leverages convolutional kernels and Fast Fourier Transform (FFT) acceleration to efficiently compute spatial co-occurrence curves across multiple radii. These curves quantify how often two features occur at defined spatial distances, enabling topographic analysis of tissue structure. Application to synthetic and biological datasets demonstrates that Sonar can recover meaningful spatial patterns such as periodic islet spacing, intra-islet cell arrangements, and spatial motifs in complex tissues. A Sonar-based generative algorithm is also presented that re-engineers cell-type maps from empirically determined co-occurrence indicators, useful for illustrating the structural information captured by spatial statistics. The ovrlpy tool identifies vertical signal contamination as critical source of noise in SRT data analysis. It defines a vertical signal integrity (VSI) score that can be used identify spatial doublets and tissue folding artefacts. These artefacts can confound downstream analyses, their removal is shown to enhance segmentation-based workflows, yielding more precise gene expression models and cell type cluster separation. Together, these tools contribute to a segmentation-free framework for spatial transcriptomic analysis, grounded in spatial statistics, scalable computation, and quality control using a noise model dedicated to spatial data. Overall, the thesis advocates for a stronger decoupling of spatial transcriptomics analysis from single-cell inspired conventions and adoption of a dedicated analytic landscape, complete with its own methods, benchmarks, and conceptual models, to fully leverage the unique spatial nature of SRT data.
