Instituto de Telecomunicações

Funder

Organizational Unit

Publications

Fine-tuning pre-trained neural networks for medical image classification in small clinical datasets

Publication . Spolaôr, Newton; Lee, Huei Diana; Mendes, Ana Isabel; Nogueira, Conceição; Parmezan, Antonio Rafael Sabino; Takaki, Weber Shoity Resende; Coy, Claudio Saddy Rodrigues; Wu, Feng Chung; Fonseca-Pinto, Rui

Convolutional neural networks have been effective in several applications, arising as a promising supporting tool in a relevant Dermatology problem: skin cancer diagnosis. However, generalizing well can be difficult when little training data is available. The fine-tuning transfer learning strategy has been employed to differentiate properly malignant from non-malignant lesions in dermoscopic images. Fine-tuning a pre-trained network allows one to classify data in the target domain, occasionally with few images, using knowledge acquired in another domain. This work proposes eight fine-tuning settings based on convolutional networks previously trained on ImageNet that can be employed mainly in limited data samples to reduce overfitting risk. They differ on the architecture, the learning rate and the number of unfrozen layer blocks. We evaluated the settings in two public datasets with 104 and 200 dermoscopic images. By finding competitive configurations in small datasets, this paper illustrates that deep learning can be effective if one has only a few dozen malignant and non-malignant lesion images to study and differentiate in Dermatology. The proposal is also flexible and potentially useful for other domains. In fact, it performed satisfactorily in an assessment conducted in a larger dataset with 746 computerized tomographic images associated with the coronavirus disease.

2023-08-31Journal article

Open access

A generic framework for optimal 2D/3D key-frame extraction driven by aggregated saliency maps

Publication . Ferreira, Lino; Cruz, Luis A. da Silva; Assunção, Pedro

This paper proposes a generic framework for extraction of key-frames from 2D or 3D video sequences, relying on a new method to compute 3D visual saliency. The framework comprises the following novel aspects that distinguish this work from previous ones: (i) the key-frame selection process is driven by an aggregated saliency map, computed from various feature maps, which in turn correspond to different visual attention models; (ii) a method for computing aggregated saliency maps in 3D video is proposed and validated using fixation density maps, obtained from ground-truth eye-tracking data; (iii) 3D video content is processed within the same framework as 2D video, by including a depth feature map into the aggregated saliency. A dynamic programming optimisation algorithm is used to find the best set of K frames that minimises the dissimilarity error (i.e., maximise similarity) between the original video shots of size and those reconstructed from the key-frames. Using different performance metrics and publicly available databases, the simulation results demonstrate that the proposed framework outperforms similar state-of-art methods and achieves comparable performance as other quite different approaches. Overall, the proposed framework is validated for a wide range of visual content and has the advantage of being independent from any specific visual saliency model or similarity metrics.

2015-11Journal article

Restricted access

Sparse least-squares prediction for intra image coding

Publication . Lucas, Luis F. R.; M. M. Rodrigues, Nuno; Pagliari, Carla L.; Silva, Eduardo A. B. da; Faria, Sergio M. M. de

This paper presents a new intra prediction method for efficient image coding, based on linear prediction and sparse representation concepts, denominated sparse least-squares prediction (SLSP). The proposed method uses a low order linear approximation model which may be built inside a predefined large causal region. The high flexibility of the SLSP filter context allows the inclusion of more significant image features into the model for better prediction results. Experiments using an implementation of the proposed method in the state-of-the-art H.265/HEVC algorithm have shown that SLSP is able to improve the coding performance, specially in the presence of complex textures, achieving higher coding gains than other existing intra linear prediction methods. © 2015 IEEE.

2015Conference paper

Open access

Performance, power and scalability analysis of HEVC interpolation filter using FPGAs

Publication . Gomez-Pulido, Juan A.; Cordeiro, Paulo J.; Assunção, Pedro

Motion compensation is the most time-consuming stage of the most recent video coding standard, and uses an interpolation filter to handle efficiently the video bitstream. When high resolutions, low power budgets and huge amount of video data are demanded, exploiting parallelism is a mandatory task. In this paper we propose an implementation of the interpolation filter using the reconfigurable hardware technology, in order to build parallel computing systems that offer a high performance, in terms of computing time and power consumption. The timing simulations and energy analysis performed on different devices show that the on-chip replication of the filter provides high speedups with regard to general purpose processors. The good experimental results motivates us to do a first approach to scalable parallel computing systems where parallelism is exploited from fine to coarse grain, multiplying the speedups obtained. In particular, we propose an on-chip multiprocessor system where filters act as coprocessors of embedded high-performance and low-power microprocessors, linked among them by point-to-point buses. This on-chip architecture can be applied to high performance computing systems based on the same reconfigurable hardware technology.

2015-10Conference paper

Restricted access

Contributions to lossless coding of medical images using minimum rate predictors

Publication . Joao M. Santos; Guarda, André; M. M. Rodrigues, Nuno; Faria, Sergio

Medical imaging compression is experiencing a growth in terms of usage and image resolution, namely in diagnostics systems that require a large set of images, like MRI or CT. Furthermore, legal and diagnosis restrictions impose the use of lossless compression and data archival for several years. These facts create a demand for more efficient compression tools, used for archiving and communication. In this work, we first evaluate the performance of traditional medical image compression algorithms against that of recent state of the art lossless image encoders. We then propose a method to improve the Minimum Rate Predictors lossless encoder, by exploiting inter picture redundancy in volumetric anatomical images. Results show that the proposed method is more efficient than state of the art encoders, such as HEVC, by about 28.8%, and achieves a gain of up to 57.8% in compression ratio when compared with traditional methods.

2015-12Conference paper

Restricted access