Work

AI-driven neuronal injury classification in preclinical HIE drug discovery; Dataset and Code

Year: 2026
Type: dataset
Abstract: This record contains the code, trained models and sample data for the paper "AI-driven neuronal injury classification in preclinical HIE drug discovery" (Loomes, Boulefred-Derrar et al., 2026).Ten deep-learning architectures (CNNs, vision transformers and CNN–transformer hybrids) were compared on classifying neurons in fetal sheep brain histology as healthy, intermediate or pyknotic. Each model takes a 4-channel input: the RGB image of a single cell plus a binary mask of that cell. Every model was trained with 5-fold cross-validation on tissue from sham, hypoxia–ischemia (HI) and HI + hypothermia animals. The models were then tested, with no fine-tuning, on a separate drug-treated (out-of-domain) test set covering six brain regions.Contents:Code: shared modules for data loading, the ten model architectures, training and evaluation. There is one training script per model, plus scripts for validation inference, external test inference and inference-speed benchmarking.Trained checkpoints for the three best-performing models: SkipNet (a lightweight CNN trained from scratch), EfficientNet-B7 and Hybrid-ViT-384. Each has five checkpoints, one per cross-validation fold.Results for all ten models: training logs, per-fold and per-class F1 scores, confusion matrices, per-region results, cell-level test predictions and inference-speed benchmarks.A Jupyter notebook that reproduces every data figure in the paper and supplementary material.The labelling tool used to crop and annotate the cells.Sample data, so the pipeline can be run end to end:36 in-domain cells, 12 per class, chosen at random;36 external test cells, 12 per class, one per class from each of the six brain regions of two animals (13-062 and 13-083). No cells from the MPA-treated animal (22-104) are included.The full dataset (13,529 in-domain cells and the complete external test set) and the checkpoints for the other seven models are available from the corresponding author on reasonable request.The code is written in Python with PyTorch, torchvision and Hugging Face Transformers. Training requires a CUDA GPU. Installation, the folder layout and step-by-step run instructions are in README.md.
Language: English
Subfield: Biophysics
Open Access status: green
OpenAlex ID: W7225257941
Cites:
Cited by:
Related to: 10