Pancreatic cancer is one of the deadliest cancers, with a high mortality
rate due to late detection and limited diagnostic accuracy. Effective and
early classification of pancreatic tumors from medical images is critical
to improving patient outcomes. Conventional deep learning methods
often struggle with overfitting, imbalanced datasets, and lack of
generalization across different image modalities. Existing single-model
deep learning approaches lack robustness and accuracy, particularly
in the classification of complex and heterogeneous pancreatic tumors.
There is a growing need for a scalable and ensemble-based solution to
enhance diagnostic accuracy while minimizing false predictions. This
study proposes an ensemble deep learning framework that integrates
three high-performing convolutional neural networks (CNNs):
ResNet50, DenseNet201, and InceptionV3. Each model is fine-tuned
on a curated pancreatic tumor dataset using transfer learning and
combined using a weighted majority voting mechanism. The
framework enhances feature extraction diversity and leverages
complementary model strengths. The proposed ensemble model
achieved superior performance over individual models and existing
hybrid approaches. Specifically, it attained an overall accuracy of
96.3%, precision of 95.1%, recall of 96.7%, F1-score of 95.9%, and
AUC of 0.982 on the test dataset. Compared to state-of-the-art hybrid
models such as CNN-SVM, ResNet-GRU, and DenseNet-LSTM, our
method demonstrated higher stability and generalization in
classification.
W. Nancy1, D. Ruban Thomas2
Jeppiaar Institute of Technology, India1, Vel Tech Multi Tech Dr. Rangarajan Dr. Sakunthala Engineering College, India2
Pancreatic Tumor, Ensemble Learning, Deep Learning, Image Classification, Medical Imaging
| January | February | March | April | May | June | July | August | September | October | November | December |
| 0 | 0 | 0 | 0 | 2 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
| Published By :
ICTACT
Published In :
ICTACT Journal on Image and Video Processing
( Volume: 15 , Issue: 4 , Pages: 3582 - 3588 )
Date of Publication :
May 2025
Page Views :
12
Full Text Views :
2
|