Wyniki wyszukiwania - BazTech

Ograniczanie wyników

Znaleziono wyników: 3

Liczba wyników na stronie

Wyniki wyszukiwania

Wyszukiwano:
w słowach kluczowych: fine-tuning

Sortuj według:

Ogranicz wyniki do:

Optimizing ultrasound image classification through transfer learning: fine-tuning strategies and classifier impact on pre-trained inner-layers

Bal-Ghaoui Mohamed, Alaoui Hachem El Yousfi, Jilbab Abdelilah, Bourouhou Abdennaser

Informatyka, Automatyka, Pomiary w Gospodarce i Ochronie Środowiska

2023

T. 13, nr 4

27--33

Transfer Learning (TL) is a popular deep learning technique used in medical image analysis, especially when data is limited. It leverages pre-trained knowledge from State-Of-The-Art (SOTA) models and applies it to specific applications through Fine-Tuning (FT). However, fine-tuning large models can be time-consuming, and determining which layers to use can be challenging. This study explores different fine-tuning strategies for five SOTA models (VGG16, VGG19, ResNet50, ResNet101, and InceptionV3) pre-trained on ImageNet. It also investigates the impact of the classifier by usinga linear SVM for classification. The experiments are performed on four open-access ultrasound datasets related to breast cancer, thyroid nodules cancer, and salivary glands cancer. Results are evaluated using a five-fold stratified cross-validation technique, and metrics like accuracy, precision, and recall are computed. The findings show that fine-tuning 15% of the last layers in ResNet50 and InceptionV3 achieves good results. Using SVM for classification further improves overall performance by 6% for the two best-performing models. This research provides insights into fine-tuning strategiesandthe importance of the classifier in transfer learning for ultrasound image classification.

Transfer Learning (TL) to popularna technika głębokiego uczenia stosowana w analizie obrazów medycznych, zwłaszcza gdy ilość danych jestograniczona. Wykorzystuje ona wstępnie wyszkoloną wiedzę z modeli State-Of-The-Art (SOTA) i zastosowanie ich do konkretnych aplikacji poprzez dostrajanie (Fine-Tuning –FT). Jednak dostrajanie dużych modeli może być czasochłonne, a określenie, których warstw użyć, może stanowić wyzwanie.W niniejszym badaniu przeanalizowano różne strategie dostrajania dla pięciu modeli SOTA (VGG16, VGG19, ResNet50, ResNet101 i InceptionV3) wstępnie wytrenowanych na ImageNet. Zbadano również wpływ klasyfikatora przy użyciu liniowej SVM do klasyfikacji. Eksperymenty przeprowadzonona czterech ogólnodostępnych zbiorach danych ultrasonograficznych związanych z rakiem piersi, rakiem guzków tarczycy i rakiemgruczołów ślinowych. Wyniki są oceniane przy użyciu techniki pięciowarstwowej walidacji krzyżowej, a wskaźniki takie jak dokładność, precyzja i odzyskiwanie są obliczane. Wyniki pokazują, że dostrojenie 15% ostatnich warstw w ResNet50 i InceptionV3 osiąga dobre wyniki. Użycie SVM do klasyfikacjidodatkowo poprawia ogólną wydajność o 6% dla dwóch najlepszych modeli. Badania te zapewniają informacje na temat strategii dostrajania i znaczenia klasyfikatoraw uczeniu transferowym dla klasyfikacji obrazów ultrasonograficznych.

Leveraging Transfer Learning to Identify Food Categories

Kolla J. V. V., Vemula P. Ch., Chakravarthy S., Naidu B. S., Patibandla D.

Advances in Science and Technology. Research Journal

2021

Vol. 15, no 4

101--109

In today’s scenario, recognition of pictured food dishes automatically has significant importance. During the COVID-19 pandemic, there was a decline in people visiting restaurants for their dietary requirements. So many restaurants started offering their services online. This situation caused a demand for better categorization of food into various categories on a large scale by companies that facilitated these services. It is challenging to congregate a large dataset of food categories, so it is complex to build a generalized architecture. To solve this issue, In this paper, domain-specific transfer learning is used to build the model using some standard architectures like VGGNET, RESNET, and EFFICIENTNET family, which are trained on popular benchmark datasets such as IMAGENET, COCO, etc. The similarity between the source and target datasets is calculated to find the best source dataset, and the one with the highest similarity is chosen for transfer learning. The solution proposed in this paper outperforms some of the existing works on categorizing food items.

Optical coherence tomography image for automatic classification of diabetic macular edema

Wang Ping, Li Jia-Li, Ding Hao

Optica Applicata

2020

Vol. 50, nr 4

567--577

Diabetic macular edema (DME) is the dominant reason of diabetic visual loss, so early detection and treatment of DME is of great significance for the treatment of diabetes. Based on transfer learning, an automatic classification method is proposed to distinguish DME images from normal images in optical coherence tomography (OCT) retinal fundus images. Features of the DME are automatically identified and extracted by the pre-trained convolutional neural network (CNN), which only involves fine-tuning the VGGNet-16 network without any user intervention. An accuracy of 97.9% and a sensitivity of 98.0% are acquired with the OCT images in the Duke data set from experimental results. The proposed method, a core part of an automated diagnosis system of the DME, revealed the ability of fine-tuning models to train non-medical images, allowing them can be classified with limited training data. Moreover, it can be developed to assist early diagnosis of the disease, effectively delaying (or avoiding) the progression of the disease, consequently.