Abstract Single-cell RNA sequencing has enabled the construction of comprehensive cell atlases, yet the quality and coherence of the cell-type annotations within these atlases remain largely unexamined. When a label is applied to a transcriptionally heterogeneous population, the downstream analyses that depend on it, and automated label transfer in particular, become unreliable. We present scINTILLA (Single-Cell Integrated Inference, Labelling, and Landscape Analysis), a computational framework that combines supervised and unsupervised machine learning to score the learnability and internal consistency of cell-type labels in single-cell datasets. The unsupervised arm benchmarks a broad panel of clustering algorithms and derives a neighbourhood confusion score for every cell, whilst the supervised arm trains up to twelve classifiers and extracts prediction agreement, entropy, and confidence. These signals are normalised and aggregated into a single composite score per cell type, where a low score flags label ambiguity or concealed heterogeneity. As a by-product, scIN-TILLA also reports which clustering and classification algorithms perform best on a given dataset, offering practical guidance for downstream label transfer. We applied it to five Human Cell Atlas datasets spanning the adult brain, lung, eye, and two organoid atlases, and recovered clear differences in the learnability and internal consistency of annotations across atlases that were not driven by the number of annotated cell types. Focused re-analysis of lowscoring populations in the lung and endoderm-organoid atlases resolved biologically coherent sub-populations, in some cases with context-specific enrichment, much of it recovered from cells that had been assigned broad or catch-all labels. scINTILLA is advisory rather than prescriptive, guiding principled, data-driven re-annotation at atlas scale.

Deep learning models have shown their potential for several applications. However, most of the models are opaque and difficult to trust due to their complex reasoning - commonly known as the black-box problem. Some fields, such as medicine, require a high degree of transparency to accept and adopt such technologies. Consequently, creating explainable/interpretable models or applying post-hoc methods on classifiers to build trust in deep learning models are required. Moreover, deep learning methods can be used for segmentation tasks, which typically require hard-to-obtain, time-consuming manually-annotated segmentation labels for training. This paper introduces three inherently-explainable classifiers to tackle both of these problems as one. The localisation heatmaps provided by the networks -- representing the models' focus areas and being used in classification decision-making -- can be directly interpreted, without requiring any post-hoc methods to derive information for model explanation. The models are trained by using the input image and only the classification labels as ground-truth in a supervised fashion - without using any information about the location of the region of interest (i.e. the segmentation labels), making the segmentation training of the models weakly-supervised through classification labels. The final segmentation is obtained by thresholding these heatmaps. The models were employed for the task of multi-class brain tumour classification using two different datasets, resulting in the best F1-score of 0.93 for the supervised classification task while securing a median Dice score of 0.67$\pm$0.08 for the weakly-supervised segmentation task. Furthermore, the obtained accuracy on a subset of tumour-only images outperformed the state-of-the-art glioma tumour grading binary classifiers with the best model achieving 98.7\% accuracy.

Biobank-scale imaging provides an unprecedented opportunity to characterise how thousands of organ phenotypes vary in populations. However, deriving specific phenotypes from imaging data requires time-consuming expert annotation, limiting scalability. In this study, we develop a 3D diffusion autoencoder to derive latent phenotypes from temporally resolved cardiac MRI data of 71,017 UK Biobank participants. These phenotypes are reproducible, heritable (h2 = [4-18%]), and significantly associated with cardiometabolic traits. To establish the genetic basis of such traits, we perform a genome-wide association study, identifying 89 significant common variants (P < 2.3 × 10-9) across 42 loci, including seven novel loci. Extensive multi-trait colocalisation analyses (PP.H4 > 0.8) link variants across phenotypic scales, from intermediate cardiac traits to cardiac disease endpoints. In conclusion, this study showcases the use of diffusion autoencoding methods as powerful tools for unsupervised phenotyping, genetic discovery and disease risk prediction using cardiac MRI data.

Liver tumour ablation presents a significant clinical challenge: whilst tumours are clearly visible on pre-operative MRI, they are often effectively invisible on intra-operative CT due to minimal contrast between pathological and healthy tissue. This work investigates the feasibility of cross-modality weak supervision for scenarios where pathology is visible in one modality (MRI) but absent in another (CT). We present a hybrid registration-segmentation framework that combines MSCGUNet for inter-modal image registration with a UNet-based segmentation module, enabling registration-assisted pseudo-label generation for CT images. Our evaluation on the CHAOS dataset demonstrates that the pipeline can successfully register and segment healthy liver anatomy, achieving a Dice score of 0.72. However, when applied to clinical data containing tumours, performance degrades substantially (Dice score of 0.16), revealing the fundamental limitations of current registration methods when the target pathology lacks corresponding visual features in the target modality. We analyse this “domain gap” and “feature absence” problem, demonstrating that whilst spatial propagation of labels via registration is feasible for visible structures, segmenting truly invisible pathology remains an open challenge. Our findings highlight that registration-based label transfer cannot compensate for the absence of discriminative features in the target modality, providing important insights for future research in cross-modality medical image analysis.

The dependence on expert annotation has long constituted the primary rate-limiting step in the application of artificial intelligence to biomedicine. While supervised learning drove the initial wave of clinical algorithms, a paradigm shift towards unsupervised and self-supervised learning (SSL) is currently unlocking the latent potential of biobank-scale datasets. By learning directly from the intrinsic structure of data - whether pixels in a magnetic resonance image (MRI), voxels in a volumetric scan, or tokens in a genomic sequence - these methods facilitate the discovery of novel phenotypes, the linkage of morphology to genetics, and the detection of anomalies without human bias. This article synthesises seminal and recent advances in "learning without labels," highlighting how unsupervised frameworks can derive heritable cardiac traits, predict spatial gene expression in histology, and detect pathologies with performance that rivals or exceeds supervised counterparts.

The lenticulostriate arteries (LSAs) supply critical subcortical brain structures and are affected in cerebral small vessel disease (CSVD). Changes in their morphology are linked to cardiovascular risk factors and may indicate early pathology. 7T Time-of-Flight MR angiography (TOF-MRA) enables clear LSA visualisation. We aimed to develop a semi-automated pipeline for quantifying 3D LSA morphology from 7T TOF-MRA in CSVD patients. We used data from a local 7T CSVD study to create a pipeline, LUMEN, comprising two stages: vessel segmentation and LSA quantification. For segmentation, we fine-tuned a deep learning model, DS6, and compared it against nnU-Net and a Frangi-filter pipeline, MSFDF. For quantification, centrelines of LSAs within basal ganglia were extracted to compute branch counts, length, tortuosity, and maximum curvature. This pipeline was applied to 69 subjects, with results compared to traditional analysis measuring LSA morphology on 2D coronal maximum intensity projection (MIP) images. For vessel segmentation, fine-tuned DS6 achieved the highest test Dice score (0.814±0.029) and sensitivity, whereas nnU-Net achieved the best balanced average Hausdorff distance and precision. Visual inspection confirmed that DS6 was most sensitive in detecting LSAs with weak signals. Across 69 subjects, the pipeline with DS6 identified 23.5 ± 8.5 LSA branches. Branch length inside the basal ganglia was 26.4 ± 3.5 mm, and tortuosity was 1.5 ± 0.1. Extracted LSA metrics from 2D MIP analysis and our 3D analysis showed fair-to-moderate correlations. Outliers highlighted the added value of 3D analysis. This open-source deep-learning-based pipeline offers a validated tool quantifying 3D LSA morphology in CSVD patients from 7T-TOF-MRA for clinical research.

Hyperthermia (HT) in combination with radio- and/or chemotherapy has become an accepted cancer treatment for distinct solid tumour entities. In HT, tumour tissue is exogenously heated to temperatures between 39 and 43 °C for 60 min. Temperature monitoring can be performed non-invasively using dynamic magnetic resonance imaging (MRI). However, the slow nature of MRI leads to motion artefacts in the images due to the movements of patients during image acquisition. By discarding parts of the data, the speed of the acquisition can be increased - known as undersampling. However, due to the invalidation of the Nyquist criterion, the acquired images might be blurry and can also produce aliasing artefacts. The aim of this work was, therefore, to reconstruct highly undersampled MR thermometry acquisitions with better resolution and with fewer artefacts compared to conventional methods. The use of deep learning in the medical field has emerged in recent times, and various studies have shown that deep learning has the potential to solve inverse problems such as MR image reconstruction. However, most of the published work only focuses on the magnitude images, while the phase images are ignored, which are fundamental requirements for MR thermometry. This work, for the first time, presents deep learning-based solutions for reconstructing undersampled MR thermometry data. Two different deep learning models have been employed here, the Fourier Primal-Dual network and the Fourier Primal-Dual UNet, to reconstruct highly undersampled complex images of MR thermometry. MR images of 44 patients with different sarcoma types who received HT treatment in combination with radiotherapy and/or chemotherapy were used in this study. The method reduced the temperature difference between the undersampled MRIs and the fully sampled MRIs from 1.3 to 0.6 °C in full volume and 0.49 °C to 0.06 °C in the tumour region for a theoretical acceleration factor of 10.