Improving convolutional neural networks performance for image classification using test time augmentation

ǂa ǂcase study using MURA dataset

Ibrahem Kandel (Author), Mauro Castelli (Author)

Abstract

Bone fractures are one of the main causes to visit the emergency room (ER); the primary method to detect bone fractures is using X-Ray images. X-Ray images require an experienced radiologist to classify them; however, an experienced radiologist is not always available in the ER. An accurate automatic X-Ray image classifier in the ER can help reduce error rates by providing an instant second opinion to the emergency doctor. Deep learning is an emerging trend in artificial intelligence, where an automatic classifier can be trained to classify musculoskeletal images. Image augmentations techniques have proven their usefulness in increasing the deep learning model's performance. Usually, in the image classification domain, the augmentation techniques are used during training the network and not during the testing phase. Test time augmentation (TTA) can increase the model prediction by providing, with a negligible computational cost, several transformations for the same image. In this paper, we investigated the effect of TTA on image classification performance on the MURA dataset. Nine different augmentation techniques were evaluated to determine their performance compared to predictions without TTA. Two ensemble techniques were assessed as well, the majority vote and the average vote. Based on our results, TTA increased classification performance significantly, especially for models with a low score.

Keywords

image classification;convolutional neural networks;transfer learning;test time 29 augmentation;deep learning;ensemble learning;

Data

Language:	English
Year of publishing:	2021
Typology:	1.01 - Original Scientific Article
Organization:	UL EF - Faculty of Economics
UDC:	659.2:004
COBISS:	72230659
ISSN:	2047-2501
Views:	331
Downloads:	29
Average score:	0 (0 votes)
Metadata:

Other data

Secondary language:	English
Type (COBISS):	Article
Embargo end date (OpenAIRE):	2022-08-01
Pages:	22 str.
Volume:	ǂVol. ǂ9
Issue:	ǂart ǂ33
Chronology:	2021
DOI:	10.1007/s13755-021-00163-7
ID:	13296237