BAAnet: an efficient deep neural network for automatic bone age assessment
Nenhuma Miniatura disponível
Data
2020-07-14
Autores
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Federal de Goiás
Resumo
The task of bone age assessment is constantly performed by radiologists worldwide to aid in the diagnosis of metabolic
and endocrine disorders in children. Currently, the most used methods to perform such task are manual methods based
on image comparison developed between the 1950s and 1960s. These methods make the evaluation process long,
tedious and costly, besides presenting high variability and uncertainty between examinations and also among
examiners. This work starts describing the characteristics of the bone age evaluation problem. It then presents some
automatic solutions proposed in the last decades. Later it describes the paradigm shift in the field of computer vision
with the development of convolutional neural networks and automatic feature extraction strategies, but also points out
the high computational cost of newly published neural networks when applied to high resolution images such as the
ones from medical imaging applications. Finally, it proposes a novel convolutional neural network to perform the task of
bone age assessment in a computationally efficient manner. This network is named BAAnet and it introduces a novel
convolutional module called the Incremental Convolutional Estimation (ICE) module. We evaluate the module against
two standard benchmarks and see an average relative improvement of approximately 10\% in performance. We also
submitted an ensemble of BAAnet models for a competition organized by the Radiological Society of North America
(RSNA) with the goal to obtain the lowest mean absolute error in a dataset of 14236 images collected for the event. Our
submission finished third and had a mean absolute error of 4.38 months (less than 3\% lower than the first place) in a
dataset of 200 images used exclusively for final evaluation of the competing models. BAAnet achieved this performance
with a number of parameters and computational cost an order of magnitude smaller than the models in first and second
places.
Descrição
Palavras-chave
Citação
PEREIRA, L. A. BAAnet: an efficient deep neural network for automatic bone age assessment. 2020. 54 f. Dissertação (Mestrado em Ciência da Computação) - Universidade Federal de Goiás, Goiânia, 2020.