Downloads · 30 days
0
ShubhamGajjar/skin-cancer-hybrid-resnet-vit
skin-cancer-hybrid-resnet-vit is a image classification model from ShubhamGajjar. Use it when you need a label for an image. The card lists the license as mit.
This model corresponds to the Hybrid ResNet50–ViT architecture from the IEEE AIC 2025 paper "A Hybrid ResNet-ViT Architecture for Skin Cancer Classification". Code: GitHub · Paper: IEEE Xplore
Downloads · 30 days
0
Access
Public
Updated Mar 7, 2026
Repo size
147 MB
Likes
0
Public
Click a slice to open those files.
.h5147 MB · 100%
From the Hugging Face model README
This model corresponds to the Hybrid ResNet50–ViT architecture from the IEEE AIC 2025 paper "A Hybrid ResNet-ViT Architecture for Skin Cancer Classification". Code: GitHub · Paper: IEEE Xplore
This repository hosts the TensorFlow/Keras weights for the model described in:
A Hybrid ResNet-ViT Architecture for Skin Cancer Classification
IEEE AIC, 2025.
Available at: https://ieeexplore.ieee.org/document/11212073
The model is trained on the HAM10000 dermatoscopic image dataset for multi-class skin lesion classification.
Abstract. Correct identification of different skin lesions, such as melanoma, from dermatoscopic images is a major challenge for automated systems, thus causing delay in early diagnosis. Most of the current models are unable to achieve an optimal balance between local feature extraction and global context understanding. Our work overcomes this challenge by proposing a hybrid ResNet50–ViT model. This architecture is significant because it allows the categorization of seven lesion classes from the HAM10000 dataset by fusing ViT's global context awareness with ResNet's strong local feature extraction. For improving its robustness, data augmentation methods were used. The designed model demonstrated a testing accuracy of 96.3% and showed better discriminative ability, as indicated through ROC values of nearly 1.00 for all test set classes. This hybrid model classifies different types of skin lesions with high accuracy, which is a major breakthrough in computer-aided dermatological diagnosis.
.h5 checkpoint).The hybrid architecture combines strong local feature extraction (ResNet) with global context modeling (ViT), which is particularly useful for complex skin lesion appearances and subtle global patterns.
Intended use
Limitations
The checkpoint super_model_v8.h5 is the one that achieved 96.3% accuracy and AUC ~ 1.00 on the HAM10000 test set and matches the model reported in the IEEE AIC 2025 paper. The GitHub repository provides a PyTorch implementation that mirrors the same architecture for research or extension; use this Keras checkpoint to reproduce the paper’s numbers or run inference in TensorFlow.
from huggingface_hub import hf_hub_download
from tensorflow import keras
repo_id = "ShubhamGajjar/skin-cancer-hybrid-resnet-vit"
ckpt_path = hf_hub_download(
repo_id=repo_id,
filename="super_model_v8.h5",
)
model = keras.models.load_model(ckpt_path)
You can then preprocess images according to your own pipeline or follow the same preprocessing described in the accompanying GitHub repository:
SkinCancer-Hybrid-ResNet-ViT (code, data pipeline, training and evaluation scripts)akiec, bcc, bkl, df, mel, nv, vasc).data.py (training_transforms / eval_transforms).Full details of the training configuration (optimizer, learning rate schedule, augmentation, epochs) are provided in the accompanying paper and code release.
On the HAM10000 test split (as defined in the paper), the model achieves:
Exact metrics, including per-class results and confusion matrix, can be reproduced using the scripts in the GitHub repository.
For reproducible experiments:
If you use this model, please cite:
A Hybrid ResNet-ViT Architecture for Skin Cancer Classification
IEEE AIC, 2025.
Available at: https://ieeexplore.ieee.org/document/11212073
You can also use the citation metadata from the CITATION.cff file in the GitHub code repository.