 Performance Analysis of Big Transfer Models on Biomedical Image ClassificationAbhishek Laxman Joshi,
Iswarya Kannoth Veetil,
B. Premjith (),
V. Sowmya,
E. A. Gopalakrishnan and
Ravi Vinayakumar
A chapter in Analytics Modeling in Reliability and Machine Learning and Its Applications, 2025, pp 141-160 from Springer Abstract: Abstract Doctors can effectively manage patients’ treatments and diseases by leveraging advanced medical imaging, which significantly minimizes guesswork and enhances diagnoses and treatments.The use of Deep Learning (DL) has been increasing recently in the area of medical imaging for various diseases like Parkinson’s, Alzheimer’s, Blood Cancer etc. When it comes to medical imaging, one common problem prevailing is that of class imbalance. There have been several proposals for the classification of Parkinson’s Disease (PD) using Machine Learning (ML) and Deep Learning (DL) techniques on an imbalanced dataset. These approaches have utilized various image pre-processing techniques on datasets such as T1-weighted and T2-weighted MRI images. One of the challenges faced by DL and ML models is the class imbalance problem in the data, where the Parkinson’s class has a majority of the data, and the normal class has a minority. To address this issue, the study implements transfer learning, a technique that improves model performance by generating better feature representations when there is limited data. While transfer learning is commonly used to enhance performance on tasks with limited data, its effectiveness on medical images, such as MRI, is not well established. However, recent studies have shown promising results. In this study, we propose a novel transfer learning approach for the classification of Parkinson’s disease using an imbalanced Parkinson’s Progression Markers Initiative (PPMI) dataset. The approach uses large scale pre-trained networks, specifically Big Transfer (BiT) models developed by Kolesnikov. BiT mainly stresses on using of Group Normalization (GN) with Weight Standardization (WS) instead of Batch Normalization (BN). The other focus is on using of BiT-HyperRule for fine tuning, where the values which have previously performed better on the natural images are used for fine-tuning for all type of datasets. The study employs different architectures of BiT models, including BiT-S and BiT-M, namely BiT-S50x1, BiT-S50x3, BiT-S101x1, BiT-S101x3, BiT-S152x4, BiT-M50x1,BiT-M50x3, BiT-M101x1, BiT-M101x3, BiT-152x4 which are publicly available. The results show that the proposed approach using BiT-M152x4 outperforms the State of Art Model, Re-weighted Adversarial Graph Convolutional network (RA-GCN), which was experimented on the imbalanced PPMI dataset. RA-GCN was giving an accuracy of 76%, whereas the BiT-M152x4 (the best performing model) was able to give an accuracy of 86.71%. As an extension to this work, we experimented with the BiT models even on the imbalanced Blood Cell Count and Detection (BCCD) dataset with the same values of hyper-parameters that was used for the PPMI dataset, and got the better results compared to existing State-Of-Art model which was 74% for VGG16, whereas the best model in our experiment which is BiT-M152x4 gave an accuracy of 98.52%. Keywords: BiT; Group normalization; Bit-HyperRule; Imbalanced dataset (search for similar items in EconPapers) There are no downloads for this item, see the EconPapers FAQ for hints about obtaining it. Related works: Export reference: BibTeX RIS (EndNote, ProCite, RefMan) HTML/Text Persistent link: https://EconPapers.repec.org/RePEc:spr:ssrchp:978-3-031-72636-1_7 Ordering information: This item can be ordered from DOI: 10.1007/978-3-031-72636-1_7 Access Statistics for this chapter More chapters in Springer Series in Reliability Engineering from Springer |
Is your work missing from RePEc? Here is how to contribute.
Questions or problems? Check the EconPapers FAQ or send mail to .
EconPapers is hosted by the
School of Business at Örebro University.