Learned image resizing with efficient training (LRET) facilitates improved performance of large-scale digital histopathology image classification models
Year: 2026
Type: article
Abstract: Training large-scale, comprehensive histopathology image classification models using deep learning is limited by two key challenges: the small input dimensions of sub-patch images exclude architectural features that are often critical to distinguish classes, and the memory requirements needed to train comprehensive models across numerous classes results in massively protracted training times. We introduce Learned Resizing with Efficient Training (LRET), a novel approach to training large-scale histopathology classification models that preserves architectural features while enabling efficient use of computational resources. We developed two LRET architectures: one using Google's learned resizing (LRET-GLR) and our novel high-dimensional feature embedding method (LRET-HFE). Both couple resizing modules with efficient training techniques to handle large input patches. We evaluated performance on three datasets with varying complexity across five deep convolutional neural network architectures, comparing against bilinear resizing and direct training without resizing. LRET methods seamlessly integrated with diverse network architectures and adapted to multiple input dimensions. On a 74-class brain tumor dataset, LRET-HFE significantly improved training efficiency while achieving 94.16% accuracy (weighted F1 score: 0.94), representing a 19.2% improvement over bilinear resizing. Visualization of deep learned feature embeddings demonstrated the model captured meaningful representations, effectively separating tumor lineages and distinguishing tumor from normal tissue. LRET methods make training large comprehensive histopathology classification models from scratch computationally feasible by improving training efficiency while maintaining high accuracy. This approach enables training of models that capture a pathologist-level field of view and span the number of tumor classes encountered in clinical practice, facilitating broader applications in computational pathology.
Source: Heliyon
Authors: Md Zahangir Alom, Quynh T. Tran, Brent A. Orr
Institution: St. Jude Children's Research Hospital
Language: English
Topic: AI in cancer detection
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,270
OpenAlex ID: W4391157463