Overfitting in Histopathology Model Training: The Need for Customized Architectures

Saghir Alfasly, Ghazal Alabtah, H.R. Tizhoosh · 2025

Read Overfitting in Histopathology Model Training: The Need for Customized Architectures by Saghir Alfasly, Ghazal Alabtah, H.R. Tizhoosh (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-16631.txt.

Description

Saghir Alfasly, Ghazal Alabtah, and Tizhoosh present a 2025 paper on overfitting in deep learning for tissue slide images. Models built for natural photos are often fine-tuned for histopathology, the study of tissue under the microscope. The team shows this habit backfires: big borrowed models memorize rather than learn and perform worse than expected. They tested many architectures, including ResNet variants and vision transformers, on a public esophageal cancer data set.

Growing model size did not reliably help on these medical images. Simpler, purpose-built designs matched or beat the large ones while overfitting less. The problem bites hardest when data sets are small, as medical ones often are. The authors argue the field needs architectures designed for tissue images from the start. Their message runs against the default of importing the biggest available model.

Catalog record

CREATORSaghir Alfasly, Ghazal Alabtah, H.R. Tizhoosh
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE25 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-16631

Browse the full library · About the Tower of Babel