Multi-Class Gastric Histopathology Classification Using a Hybrid Swin Transformer–ConvNeXt Model

Abstract Diagnosing gastric cancer still depends, for the most part, on a pathologist looking at stained tissue slides under a microscope. The disease kills roughly 660{,}000 people every year, and although early detection makes a real difference to survival, manual slide reading is slow and two experts examining the same slide do not always agree. In this work we built a system that tries to ease that burden. We trained a hybrid deep learning model on the Gastric Cancer Histopathology Tissue Image Dataset (GCHTID), pairing a Swin Transformer, which is good at picking up the broad structural layout of a tissue patch, with a ConvNeXt network, which captures the fine local textures where early malignant changes tend to hide. Every image first passes through a simple preprocessing chain of median filtering and CLAHE contrast enhancement, a combination we settled on after comparing several denoising and enhancement options. The ensemble sorts each patch into one of eight classes: tumor tissue (TUM) plus seven benign types-adipose, debris, lymphocytes, mucus, smooth muscle, normal mucosa, and stroma. On held-out data it reached 75.35% test accuracy (88.76% training, 76.99% best validation), beating the single-model baselines we tried, including ViT and EfficientNet-B3. To make the model usable outside a notebook, we wrapped it in a Flask REST API and a small web interface where a user can upload a slide image, see the prediction and its confidence, download a PDF report, and look back through past results. The system is a research prototype rather than a clinical tool, but the results suggest that combining transformer and convolutional features is a workable direction for computer-assisted pathology.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-12
DOI
https://doi.org/10.21203/rs.3.rs-11703260/v1
Primary Topic
AI in cancer detection
Type
preprint

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Multi-Class Gastric Histopathology Classification Using a Hybrid Swin Transformer–ConvNeXt Model

Rana Fayyaz Ahmad, Muhammad Junaid Asif, Muhammad Ahsan Aziz, Usman Nazakat et al.
Research Square
AI in cancer detection
preprint

Multi-Class Gastric Histopathology Classification Using a Hybrid Swin Transformer–ConvNeXt Model

Rana Fayyaz Ahmad, Muhammad Junaid Asif, Muhammad Ahsan Aziz, Usman Nazakat, Uzair Khan, Muhammad Saad
preprint en

Abstract

Abstract Diagnosing gastric cancer still depends, for the most part, on a pathologist looking at stained tissue slides under a microscope. The disease kills roughly 660{,}000 people every year, and although early detection makes a real difference to survival, manual slide reading is slow and two experts examining the same slide do not always agree. In this work we built a system that tries to ease that burden. We trained a hybrid deep learning model on the Gastric Cancer Histopathology Tissue Image Dataset (GCHTID), pairing a Swin Transformer, which is good at picking up the broad structural layout of a tissue patch, with a ConvNeXt network, which captures the fine local textures where early malignant changes tend to hide. Every image first passes through a simple preprocessing chain of median filtering and CLAHE contrast enhancement, a combination we settled on after comparing several denoising and enhancement options. The ensemble sorts each patch into one of eight classes: tumor tissue (TUM) plus seven benign types-adipose, debris, lymphocytes, mucus, smooth muscle, normal mucosa, and stroma. On held-out data it reached 75.35% test accuracy (88.76% training, 76.99% best validation), beating the single-model baselines we tried, including ViT and EfficientNet-B3. To make the model usable outside a notebook, we wrapped it in a Flask REST API and a small web interface where a user can upload a slide image, see the prediction and its confidence, download a PDF report, and look back through past results. The system is a research prototype rather than a clinical tool, but the results suggest that combining transformer and convolutional features is a workable direction for computer-assisted pathology.

Research Square
Quaid-i-Azam University (PK), Centre for Artificial Intelligence and Robotics (IN), Artificial Intelligence in Medicine (Canada) (CA), Pak-Austria Fachhochschule: Institute of Applied Sciences and Technology (PK)
National Centre for Physics
Good health and well-being
AI in cancer detection
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon