Background
A heterozygous germline loss-of-function DICER1 variant predisposes an individual to a spectrum of solid tumours. Most associated neoplasms form after a somatic missense substitution occurs in a hotspot codon on the remaining wild-type allele. The penetrance of the predisposition is variable: many of those affected never develop a tumour whilst others develop several. What determines an individual’s phenotype is unclear.

Purpose
We present the case of a 14-year-old girl with DICER1 tumour predisposition who developed an unclassifiable pelvic malignant teratoid neoplasm after exposure to platinum chemotherapy (cisplatin) for a pineoblastoma. Platinum agents are known to leave specific mutational imprints on normal and neoplastic cells. We hypothesised that the pineoblastoma treatment caused the malignant teratoid neoplasm.

Methods
Targeted next-generation sequencing and paired tumour-normal whole genome sequencing (WGS) data were generated as part of the clinical work-up. The WGS data was re-analysed during this study. The substitutions were grouped by their clonality, using a binomial mixture model, and subjected to single base substitution (SBS) mutational signature extraction. Susceptibility to platinum-related variants across the DICER1 gene was calculated using data from the “Catalogue Of Somatic Mutations In Cancer” (COSMIC) database.

Results
The initiating DICER1 variant found within the malignant teratoid neoplasm (p.G1809R) was at a genomic locus especially susceptible to platinum-associated mutagenesis. Mutational signature analysis attributed most tumour variants to the effects of platinum (SBS31 and SBS35). There was a 93% probability that cisplatin generated the somatic DICER1 variant.

Conclusions
We have linked the genotoxic effects of chemotherapy to the tumour-initiating variant in a paediatric solid tumour. It highlights how cancer risk is shaped by the interplay between mutagen and germline variation, perhaps accounting for some of the tumour multiplicity seen in DICER1 tumour predisposition. This phenomenon is likely to be present in, and inform the management of, other cancer predisposition syndromes.

Background: Artificial Intelligence (AI) has shown great promise in diagnosing tumours in histopathological biopsy specimens. However, tumours are rare in the duodenum, limiting the availability of large datasets typically used by AI models, thus requiring data-efficient pipelines. A further challenge AI pipelines face in digital pathology is generalisation across institutions, which remains critical for clinical deployment.

Purpose: We sought to take a data-efficient approach to the development of AI for the detection of duodenal neoplasia, using data from multiple centres to ensure model generalisation.

Methods: We assembled a four-hospital dataset of haematoxylin and eosin–stained duodenal biopsies acquired at 40× magnification, comprising 343 normal and 343 tumour biopsies (including adenomas, carcinomas and neuroendocrine tumours), along with the clinical diagnosis. Whole-slide images were divided into patches, each encoded using the Hibou-b foundation model. Feature standardisation was performed using z-score normalisation, with scaling parameters estimated in an unsupervised manner. Our pipeline included principal component analysis, a Gaussian mixture (GM) preprocessing step and finally a Random Forest Classifier. Generalisability was assessed using leave-one-hospital-out cross-validation with hospital-specific feature scaling.

Results: Our model achieved 96.2% sensitivity and 83% specificity, with an average AUC of 0.96 across held-out hospitals, showing strong generalisation performance. Ablation studies demonstrated the importance of unsupervised hospital-specific standardisation (AUC increase of 0.07) and of the proposed GM pipeline (14% higher sensitivity and 4.5% higher specificity compared to mean pooling across patch embeddings).

Conclusions: We present the first AI-based approach for detecting tumours in duodenal biopsy specimens. Our results demonstrate that strong performance can be achieved using limited training data, a consideration particularly relevant given the rarity of duodenal tumours. Higher sensitivity than specificity is important in this clinical context.

Background
Mitotic count on haematoxylin and eosin-stained tumour sections is a key component of tumour grading and prognostic assessment. However, manual counting on digital whole-slide images (WSIs) is time-consuming and subject to inter-observer variability, often related to differences in hotspot selection and mitotic figure identification, which can affect diagnostic consistency and clinical decision-making. Artificial intelligence (AI) offers potential to support standardisation and reduce observer-dependent variation.

Purpose
This study aimed to evaluate whether AI-assisted mitotic counting, using MitPro, improves inter-observer consistency and reporting efficiency in routine pathology practice across multiple tumour types, institutions and systems.

Methods
We performed a two-stage, clinical validation study using WSIs from 9 different tumour types from the National Centre of Pathology, Lithuania, and North Tees and Hartlepool NHS Foundation Trust, UK.

Five validation sets were analysed: one 270-slide set and four 72-slide sets, assessed by 15 different pathologists across HALO AP and Sectra image management systems. Conventional counting was followed by AI-assisted counting after a two-week washout. The primary endpoint was inter-observer consistency, assessed using ICC(2,k), while efficiency was evaluated by time per case.

Results
ICC(2,k) increased from 0.83 to 0.99 in the 270-slide Vilnius dataset (3 pathologists, HALO AP); from 0.93 to 1.00 in the 72-slide Vilnius dataset (7 pathologists, HALO AP); from 0.83 to 0.97 in the 72-slide Vilnius dataset (5 pathologists, Sectra); from 0.95 to 0.99 in the 72-slide North Tees dataset (5 pathologists, HALO AP); and from 0.92 to 0.99 in the 72-slide North Tees dataset (5 pathologists, Sectra). Mean reporting time decreased by approximately 200 seconds per case across datasets, representing significant gains in efficiency.

Conclusion
AI-assisted mitotic counting improved inter-observer consistency and reduced reporting time across tumour types and platforms, enhancing grading reproducibility and workflow efficiency in routine histopathology.

Introduction
Diagnosing coeliac disease (CD) on duodenal biopsies is challenging, with a 20% inter-pathologist disagreement rate. Deep learning approaches promise to improve accuracy of CD diagnosis, but lack interpretability, limiting clinical adoption. We present a novel method extracting clinically meaningful morphological features, enabling interpretable diagnosis.

Purpose
We sought to complement our AI-based approach to diagnosing coeliac disease on duodenal biopsies with an interpretable deep learning approach based on morphological feature extraction.

Methods
We developed a three-stage pipeline: (1) semantic segmentation of intraepithelial lymphocytes (IELs), enterocytes, crypts, and villi, followed by (2) instance-level morphometry, extracting crypt orientation and dimensions from segmentation masks and estimating villus lengths using a trained polyline-based model, (3) a diagnostic classifier (CD-positive versus CD-negative) combining morphological measurements (villus lengths, crypt dimensions) and cellular features (IEL:enterocyte ratios). Our training data comprised: (1) 49 patches with annotated IEL, enterocyte, crypt and villus outlines, (2) 36 whole slide images (WSIs) containing 3,354 villus polyline annotations, (3) 559 WSIs with clinical diagnosis (239 CD, 236 normal and 84 CD-negative inflammation) from 4 centres.

Results
Stratified 3-fold cross-validation on CD versus normal samples achieved: accuracy 89.7% ± 3.8%, F1 score 89.7% ± 3.6%, and AUC 0.970 ± 0.036. A model trained on CD versus non-CD controls (normal and inflammatory) achieved accuracy 84.8%, F1 81.6%, AUC 0.922. External validation on data from an independent centre (162 CD, 74 normal) showed strong generalisation performance: accuracy 89.6%, F1 score 85.1%, AUC 0.967.

Conclusions
This interpretable deep learning system achieves robust diagnostic performance on data derived from multiple centres while providing clinically meaningful morphological measurements including villus and crypt dimensions, crypt orientation, and IEL:enterocyte ratios. This work demonstrates potential for more standardised, reproducible coeliac disease diagnosis.