Graphical user interface for artificial neural networks in medical imaging.

Luculescu, Marius Cristian ; Lache, Simona

1. INTRODUCTION

Computer-Aided Diagnosis (CADx) in medical imaging is a highly important inter- and transdisciplinary topic; it combines aspects from biosystems (human body), image acquisition and processing (medical imaging), artificial intelligence techniques (neural networks) and information management (databases).

By using CADx, an increasing in the initial and evolutional identification precision of disease is obtained, together with short diagnosis time, accuracy, consistence, a high confidence coefficient in results interpreting and an improvement in the prevention and treatment management.

The huge number of advantages offered by this type of diagnosis and the challenge regarding top researches on image acquisition, processing, classification/recognition using artificial intelligence methods, put together in the benefit of the human health care, have lead to such a study.

2. PROBLEM FORMULATION

2.1 Problem statement

Images of some diseases have a great variety of shapes and textures and sometimes they are hard to be identified and recognized even by experienced doctors. In these situations, a CADx system becomes a very useful tool not only for doctors, but also for researchers. That is why we are trying to develop and optimize such a system based on intelligent image classification/recognition algorithms and validating it on diagnosis of human visual system diseases, namely the macular ones.

2.2 The actual knowledge stage in the domain

The analysis of the present stage showed an extremely important interest, the computer-aided diagnosis based on image recognition being one of the major research topics within the framework of the medical imaging (Erickson & Bartholmai, 2002); (Summers, 2003). This interest is motivated by the care shown towards the people's health state, the increasing of the medical services quality and accessibility, the cost reduction of long-term treatment and hospitalization, of extra tests and of the number of surgical treatments by an early prevention or discovery of diseases. A minute analysis of the specialty literature stated that in the field of medical imaging, there have been carried out numberless research studies concerning the computer-aided diagnosis of breast cancer (Hall, 2007), of the pulmonary nodules from breast X-ray pictures (Shiraishi et al., 2003) and CAT SCANs (Lawler et al., 2003), of the cerebral aneurysms (Yokoyama et al., 2007), some areas being better covered, other areas less covered, just like the case of retinal imaging.

The in-depth analysis carried out has permitted the highlighting of certain unexplored areas towards which was oriented the research (e.g., the elaboration of some parametric identification methods in order to cover a wide range of diseases, the use of texture descriptors for images analysis a.s.o.) and of some elements that can be improved (the sensibility and specificity values as parameters defining the system performance in detecting the possibly affected areas and the recognition of the disease present in that particular area).

3. PROBLEM SOLUTION

3.1 System description

The structure of the CADx process used for macular diseases classification/recognition is presented in Fig.1. (Luculescu & Lache, 2008).

The image of the retina is acquired using a retinal investigation camera and transferred to computer. An image processing module is used for different image quality improvements or geometrical transformations.

[FIGURE 1 OMITTED]

A certain vector of image features (seven 2D moment invariants and six values that describe a region by quantifying its texture content (Gonzales et al., 2004)) is then automatically extracted and the obtained values are applied to the inputs of the artificial neural network (ANN) used for image classification/recognition. The ANN generates a response that represents a set of four possible diagnoses, ordered by the level of certitude. The doctor can establish the final diagnosis after a discussion with the patient, taking into account personal and heredo-collateral antecedents. The results are stored in a database, as a patient medical record, together with personal information and recommended treatment.

In the design phase of the system, the image recognition module using ANN has as main objectives to find out what network structure, training algorithm and training parameters are most suitable for this kind of application. That is why we have developed in Matlab[R] a parametric GUI for configuring, training, testing and optimizing the ANN.

3.2 ANN configuration

The used ANN is a feedforward total connected multilayer perceptron network type with 3 or 4 layers of neurons (1 or 2 hidden layers). The number of inputs is determined by the number of values representing features and it can be selected from the main window of the GUI. Using RGB images, for example, and seven 2D moment invariants plus six texture descriptors, the number of inputs is 39 (13 for each color component). The structure of the network can be modified by selecting one or two hidden layers and choosing the number of neurons in each layer. The number of diagnostics from the database will determine the number of outputs.

The transfer function can be selected from a list that contains tansig--hyperbolic tangent sigmoid transfer function, logsig--log sigmoid transfer function and purelin--linear transfer function. Because the teacher output vector is composed from 0 and 1 values (1 for selecting the correct diagnosis), logsig and tansig functions will be preferred considering the interval in which they are taking values.

The training parameters and the stopping conditions can be also modified from the GUI.

3.3 ANN training and testing

Training process started for a database containing 25 images of 15 diagnostics. The process can be stopped manually by user or automatically if a stop condition is fulfilled. At every stop the ANN is tested and a report regarding the number of classified diagnostics, training time and square mean error is displayed. The network can also be tested on a specific database containing images that were not used in the training phase. For example, all 25 images of the initial database were classified in less than 1000 epochs.

Due to the complexity of the shapes, distribution and texture of the images representing macular diseases it was necessary to test different training algorithms.

3.4 ANN optimization

In order to obtain the best results, an optimization is needed, regarding the structure of the network, the training algorithms and parameters.

Optimization criteria take into account maximizing the number of classified images (first priority--main objective), minimizing the training time (second priority) and minimizing the square mean error (third priority).

Restrictions refer to: the ANN type--total connected feedforward multilayer perceptron network; image database (training set)--25 RGB images, representing 15 diagnostics; ANN architecture--39-X-15 with 3 layers of neurons, where X is the variable number of neurons in the hidden layer; training process--15 cycles with 30.000 epochs each, results being stored at each 1000 epochs.

As optimization parameters were considered the training function, the transfer function and the ANN architecture.

Twelve training function were pretested, six of them being finally used for analysis: Traingd--Gradient descent backpropagation, Traingdm--Gradient descent with momentum backpropagation, Traingda--Gradient descent with adaptive learning rate backpropagation, Traingdx--Gradient descent with momentum and adaptive learning rate backpropagation, Trainrp--Resilient backpropagation and Trainscg--Scaled conjugate gradient backpropagation. Three transfer function were tested: tansig--bipolar sigmoid, logsig--unipolar sigmoid, purelin--linear function. Three architectures were used, with 10, 39, 78 neurons in the hidden layer.

3.5 Results

Best results were obtained for the trainscg training function, tansig transfer function and a 39-78-15 architecture (39 neurons in the input layer, 78 in the hidden layer and 15 in the output layer), even the training time is higher.

Then, more images were appended to the database and the training process continued cumulative until all the diagnostics were classified. Finally, the network was tested with a set of images some of them representing similar diagnostics or normal macula status. The specificity obtained was 90,91%.

Regarding performance, the percentage of images correctly detected as having potential diseases (sensitivity) and the percentage of diseases images correctly recognized (specificity) were better in our system comparing to others (Luculescu & Lache, 2008).

4. CONCLUSION

The developed GUI is a very useful tool in CADx research that allows designing and optimizing amazing image classification/recognition systems based on artificial intelligence techniques. As future research, new development directions will focus on system improvement by adding new image features that can be used as inputs in the neural network and adapting the system for recognizing images from other domains (industrial for example).

5. REFERENCES

Erickson, B.J. & Bartholmai B. (2002). Computer-Aided Detection and Diagnosis at the Start of the Third Millennium. Journal of Digital Imaging, No. 15, pp. 59-68.

Gonzales, R. et al. (2004). Digital Image Processing Using Matlab, Pearson Prentice Hall, ISBN 0-13-008519-7.

Hall, F. M. (2007). Breast Imaging and Computer-Aided Detection. The New England Journal of Medicine, Vol. 356, pp. 1464-1466.

Lawler, L.P. et al. (2003). Computer Assisted Detection of Pulmonary Nodules: Preliminary Observations Using a Prototype System with Multidetector Row CT Data Sets. Journal of Digital Imaging, No. 16, pp. 251-261.

Luculescu, M. & Lache, S. (2008). Computer-Aided Diagnosis System for Retinal Diseases in Medical Imaging, WSEAS TRANSACTIONS on SYSTEMS Journal, Issue 3, Vol. 7, March 2008, pp.264-276, ISSN 1109-2777.

Shiraishi, J. et al. (2003). Computer-Aided Diagnosis for Distinction Between Benign and Malignant Solitary Pulmonary Nodules In Chest Radiographs: ROC Analysis of Radiologists' Perform, Radiology, No. 227, pp. 469--474.

Summers, R.M. (2003). Road Maps for Advancement of Radiologic Computer-Aided Detection in The 21st Century. Radiology, No. 229, pp. 11-13.

Yokoyama, R. et al. (2007). Development of an Automated Method for the Detection of Chronic Lacunar Infarct Regions in Brain MR Images. IEICE Trans. on Inf.and Systems, No. E90-D, pp. 943-954.